0.75 g: physicists dropped antimatter to check it doesn't fall up

Here is a question that sounds silly: does antimatter fall down?

Antimatter is ordinary matter's mirror twin. Every particle has an anti-version with the same mass and the opposite charge: the electron has the positron, the proton has the antiproton. When a particle meets its twin, the two vanish in a flash of energy.

Everything you have ever dropped has fallen down. But until 2023 nobody had dropped antimatter and seen which way it fell.

The answer is down. The number is 0.75 g, give or take about a fifth. Here g is how quickly anything dropped near the ground speeds up, and ordinary matter falls at exactly 1 g. So 0.75, with its error bar, fits ordinary gravity.

By the end of this story you'll know:

how the drop was done,
why physicists were fairly sure of the answer already,
and what they are hunting now: not "up or down?" but "exactly 1 g, or a tiny bit different?"

Part oneHow the drop was done

Entry 01

It fell down

In 2023 a team at CERN, Europe's particle physics laboratory near Geneva, did something nobody had done before. The ALPHA team makes atoms of antihydrogen: each one an antiproton with a positron bound to it, the simplest anti-atom there is.

WhoThe ALPHA team, with Jeffrey Hangst as its spokesperson, made the measurement.

In a machine called ALPHA-g, they held about a hundred of these atoms at a time in an upright magnetic bottle. They let about half out sideways to count them, then opened both ends together, slowly, and counted which way the rest left. More left through the bottom.

On 27 September 2023 they reported the result in the journal Nature: antihydrogen falls towards the Earth at 0.75 g. The measurement is uncertain by about 0.21 g either way, about a fifth of g.

That sounds like "antimatter falls slower". It doesn't mean that. The gap between 0.75 and 1 is 0.25 g, about the size of the error bar itself. So the result is consistent with ordinary gravity: allowing for its uncertainty, it agrees with 1 g.

Where 0.75 g sits

0.75 ggive or take about a fifth

ruled outantigravity at full strength

ordinary gravityhow ordinary matter falls

The shaded bar is the error bar, about a fifth of g either way. The gap from 0.75 to 1 is about the same size.

What it does rule out is antigravity as strong as ordinary gravity. The team puts the chance that its data fit antimatter falling up that hard at less than one in a million billion.

It was the first direct observation of gravity acting on antimatter.

Almost no physicist was surprised. For more than sixty years there had been good reasons to expect "down". This story is about those reasons too, and why the team checked anyway.

Entry 02

Nobody had weighed it

Antimatter began as an equation. Around 1930 the physicist Paul Dirac predicted an "anti-electron" (Nº 11 meets his equation inside a car battery), and in 1932 Carl Anderson found it in cosmic rays (fast particles that rain down from space): the positron. The antiproton followed in 1955.

For about ninety years after the positron was found, nobody saw antimatter fall.

We did know that antimatter has ordinary, positive mass. When a positron meets an electron, the two vanish into flashes of light that carry exactly the energy locked up in their mass (Nº 0). And in 2022 an experiment called BASE found that an antiproton carries the same charge for its mass as a proton (with the opposite sign), to 16 parts in a trillion.

Both of those measure how hard antimatter is to push. How hard gravity pulls on it was the open question: not negative mass, which nothing ever hinted at, but whether gravity treats antimatter differently.

Why not just weigh a positron? Because it is charged, and the faintest stray electric field pushes it far harder than gravity does. In 1967 two physicists dropped electrons down a shielded tube and showed how hard that is. As far as we can find, no charged antiparticle has ever been weighed directly.

So the team used whole anti-atoms, which have no charge. Antihydrogen was first made, fast-moving, in 1995, then made cold in 2002, and ALPHA first trapped it in 2010. In 2013 ALPHA used trapped atoms to set a first, loose limit: antimatter's weight could still have been anything from 65 times normal pulling up to 110 times normal pulling down.

From an equation to a drop

c. 1930Dirac predicts an "anti-electron"
1932the positron turns up in cosmic rays
1955the antiproton
1967electrons dropped down a shielded tube
1995antihydrogen, fast-moving
2002antihydrogen, cold
2010ALPHA traps it
2013a first, loose limit
2023the drop

Entry 03

A magnetic bottle

Antihydrogen can't touch anything. One brush with ordinary matter and it annihilates, twin meeting twin. So it is held in a bottle made of magnetic fields.

The bottle stands upright, 25.6 cm tall between its lids, a bit shorter than a 30 cm school ruler. Each anti-atom is a tiny magnet, and, pointing the right way, it is pushed away from strong fields. Coils round the sides keep it off the walls; two more coils, one at the bottom and one at the top, make the lids.

Against those magnetic walls, gravity is tiny. The walls are about 1,600 times deeper than the energy an atom gains by falling the whole length of the bottle (this page's arithmetic; it compares energies, not forces).

Put another way: the most energetic atoms the bottle can hold could coast upwards about 420 m against gravity alone. Even an atom with a fiftieth of that energy could climb about 8 m. So the atoms aren't resting at the bottom: they bounce from lid to lid dozens of times a second.

How high could they coast?

Climbing against gravity alone (this page's arithmetic). Each line up the scale is ten times higher.

How can gravity show at all, then? Every atom feels it, all the time, as the same small tilt: the bottom lid is very slightly lower than the top one. The trick is to lower both lids together, slowly. Each atom leaves over whichever lid it clears first, and gravity gives the bottom lid a head start.

The head start is small, so a slow atom on its way up can still reach the top first. Gravity sets the odds rather than deciding each atom: in the team's simulation of ordinary atoms, about four in five leave by the bottom. It's like testing a loaded coin: one toss tells you nothing, many tosses show the bias. Gravity is a whisper, but every atom hears it.

Entry 04

Uncorking both ends

One try, step by step

Fill the bottleEach try began by gathering about 100 anti-atoms in the bottle, which took about four hours.
Count halfFirst, one of the sets of coils round the sides was switched off, letting out the atoms moving fastest sideways: about half of them. Counting those tells the team how many atoms it had.
Lower both lidsThen came the real release. The top and bottom lids were lowered together, steadily, over 20 seconds. Almost nothing came out in the first ten seconds. Then the other 50 or so atoms (this page's estimate) began to slip out.
Watch for flashesEach one annihilated on the wall, and a detector round the bottle saw where, so the team knew which end it had left by.
Do it again, and againThe detector catches about two in three. On average, about 18 to 32 atoms per try were counted while the lids came down; a few more leaked out afterwards and weren't used. That's far too few to read the split from one try, so the team repeated it dozens of times, 1,722 counted atoms in all.

They also tilted the bottle on purpose, with a small magnetic difference between the lids. The tilt is measured in g: a tilt of 1 g pushes as hard as gravity itself. With ordinary gravity, a tilt of −1 g cancels it, and the team's simulation puts the even split near there; antigravity would put it near +1 g. The tilts ran from −3 to +3 g, plus two extreme ones, at ±10 g, to check the detector.

Pick a tilt and a pull, then open the lids 👇

With no tilt, the real experiment saw about 7 in 10 atoms leave by the bottom, while the team's simulation of ordinary atoms gives about 8 in 10. That gap isn't, by itself, a sign of weaker gravity. The counts are small, the team's check on how evenly the detector's top and bottom see is itself uncertain, and the real tilts differ slightly from the planned ones.

Entry 05

Give or take a fifth

Guess first

The error bar is about a fifth of g. What is its biggest piece?

See the answer

The magnetic field. The biggest piece, about 0.16 g, is how well the team knows the bottle's magnetic field; counting the atoms adds only about 0.06 g.

You might guess the error bar comes from having too few atoms. Mostly, it doesn't.

The team lists its uncertainty in five pieces. The biggest, about 0.16 g, is how well it knows the bottle's magnetic field, especially away from the middle. Next, about 0.12 g, is how evenly the detector sees atoms at the top and at the bottom.

The team checked the detector with its runs at the two extreme tilts. Those counted only one or two hundred atoms at each end, so part of this piece may be counting luck too; the paper doesn't say how much. Counting the atoms in the main runs, the luck of the draw, adds only about 0.06 g, and two small pieces a little more.

The team groups the five into two error bars, about 0.13 g and 0.16 g. Separate, unrelated sources of error don't simply add up: they partly offset each other, so together they make about 0.21 g (this page's combination of the team's two error bars), not the 0.29 you'd get by simply adding them.

That matters. Use counting luck alone, and the gap from 0.75 to 1 would look like a real difference. With the full error bar, more than three times bigger, it doesn't.

Turn the knobs 👇 and watch what shrinks the total.

More atoms in the main runs alone can't get there. Even a thousand times more would leave the total near a fifth of g, because the biggest piece, the magnetic field, stays; even with more atoms in those checking runs too, it couldn't drop below about a sixth. A better field map helps most. To reach 1% of g, the goal of the next experiments, the error bar has to shrink about twenty times, and every piece has to shrink with it.

So what does 0.75 g mean? Not "antimatter falls three quarters as fast". It's the middle of a measurement uncertain by about a fifth of g. Ordinary gravity sits inside twice that uncertainty: the band from 0.34 to 1.16 g.

Part twoWhy physicists were fairly sure already

Entry 06

A machine made of light

Why were physicists so sure of "down"? The oldest reason came from Philip Morrison in 1958, and it's about energy.

Suppose antimatter fell up. Take a particle and its antiparticle together and lift them up a tower. The matter one pulls down, the antimatter one pulls up just as hard, so the lift costs nothing.

At the top, let them annihilate into light and send the light down the tower. Light gains energy as it falls, so at the bottom it has a little more than it started with: enough to make the pair again, with energy to spare. Repeat for ever, and you have a perpetual-motion machine, energy from nothing.

Morrison's tower, if antimatter fell up

1Lift the pair. If antimatter fell up, the lift costs nothing.
2At the top, annihilate it into light.
3The light falls, gaining energy.
4At the bottom, make the pair again, with energy to spare. Repeat for ever.

The argument bites because light really does fall. In 1960 two physicists measured light gaining energy down a 22.5 m tower; in 2022 a lab measured the same effect across a millimetre (Nº 9). And light is its own antiparticle, so it can't fall one way for matter and another way for antimatter.

Run the same sums for any value, and only exactly 1 g balances the books. Below it, every cycle leaves energy over; above it, the cycle run backwards does. So any difference from 1 g, not just "up", would make a machine, unless something else carries the energy (this page's sums; other physicists made the same point long ago). Try the ledger 👇

That "unless" is the loophole. In 1991 Michael Martin Nieto and Terry Goldman argued that Morrison's sums hold in Einstein's theory as it stands, but can fail if gravity has partners: a pair of new forces. On ordinary matter one would pull and the other push, so they could nearly cancel; on antimatter both would pull. Such a pair could balance the books. In the best-known idea of this kind, antimatter would fall slightly faster, not slower.

So Morrison didn't prove antigravity impossible. He showed that it would break energy conservation in Einstein's theory, and so where a loophole would have to be.

Entry 07

The antimatter inside every atom

The second reason came from Leonard Schiff in 1958 and 1959. In a way, it says there is antimatter in everything already.

NoteOther physicists also questioned Schiff's own sums.

Quantum physics says the electric field round every atom's nucleus fizzes with short-lived "virtual" pairs of electrons and positrons. They aren't real positrons sitting in the atom, but they change its energy a tiny bit, and so its mass. How much depends on the element.

If that antimatter part fell up, different materials would fall at different speeds. Careful experiments see no such difference. Twisting balances in the lab find two metals falling alike to about a part in ten million million; the MICROSCOPE satellite, to a few parts in a million billion (Nº 1).

In 2009 three physicists put numbers on Schiff's idea, in a preprint that, as far as we can find, never appeared in a journal. They assumed that any difference for antimatter shows up equally in these virtual pairs.

1 in 10 million

Under that assumption, the balances limit it to about one part in ten million. Rerun with the satellite's result, the same sum gives about one to three parts in a billion (this page's arithmetic, not a published number).

Every one of these numbers carries the same "if". Experiments on ordinary matter haven't shown that antimatter falls normally to one part in ten million. They've shown that any difference hiding in the virtual antimatter of atoms, in the same proportion as in real antimatter, must be smaller than that.

Entry 08

Kaons, clocks and a supernova

Three more clues pointed the same way.

Kaons

In 1961 Myron Good looked at particles called neutral kaons, which keep turning into their own antiparticles and back. If gravity pulled the two oppositely, their energies would split apart by far more than the tiny difference that sets the rhythm of the flipping, and kaons would behave quite differently from what we see.

The "if"The weak spot is which gravity you count: the Earth's, the Sun's or the Galaxy's give answers a thousandfold apart. Restated without that choice, the argument is much weaker.

Clocks

A trapped antiproton circles at a steady rate, so it can be used as a clock. BASE compared it with a stand-in for a proton for a year and a half, as the Earth moved nearer to the Sun and further away. The two clocks responded to the Sun's gravity alike, to within about 3%.

The "if"BASE timed the antiproton; it didn't weigh it, and the link to falling runs through energy conservation, as in chapter 6.

A supernova

In 1987 an exploding star sent us a burst of particles, mostly antineutrinos (the antimatter versions of neutrinos, ghostly particles that pass through almost anything), along with its light. On the way, our Galaxy's gravity held both back by the same amount, to within a few tenths of a per cent (Nº 7).

The "if"But Einstein's theory doesn't, on its own, say how big that hold-up is, so this test is best read as an illustration.

All three point to "down", and each carries its own assumption.

Entry 09

Why test it anyway?

With all that, why bother with a drop? The ALPHA paper itself says where opinion stood: almost every theorist agreed that anything in a lab must be attracted by the Earth.

But agreement isn't a measurement. In 1991 Nieto and Goldman went through all the old arguments and concluded that a direct test was still needed. Their worry: the hidden forces of chapter 6 could nearly cancel for ordinary matter, so the sharper the matter-only tests got, the more room they could leave for antimatter to be different.

In 2008 three physicists, asked by a committee at Fermilab, an American particle physics lab, wrote in a note (not checked by a journal) that the indirect reasoning rests on assumptions worth testing, and that nothing compelling ruled out a 1% difference. They granted, though, that the indirect evidence suggests 1 g to far better than 1%. The 2009 preprint's authors answered that even nearly cancelling forces are squeezed hard, by pairs of neutron stars (the collapsed cores of dead stars) slowly spiralling in towards each other.

A few theorists went further, and built ideas about the whole universe on antimatter being pushed away by matter. ALPHA-g ruled out the simplest versions, and some have since revised their models.

That's the sense of this story's title.

"Up" was always a long shot: almost nobody expected it, and the arguments against it were strong.

But it was never ruled out, because each argument rested on an assumption. The drop settles up versus down at the strength of ordinary gravity. It leaves small differences open.

Entry 10

Three mirrors and an apple

Physicists describe matter and antimatter with three mirrors. The first swaps every particle for its antiparticle. The second mirrors space, left for right. The third runs the film of time backwards.

Each mirror on its own is cracked: the weak force, the one behind some kinds of radioactivity, behaves differently in each mirrored world. But use all three at once and, as far as anyone has seen, the laws of physics stay exactly the same. That's the CPT rule (C for charge, P for the mirror in space, T for time), and a theorem from the 1950s says any theory of the usual kind must keep it. Antihydrogen's light, for one, matches hydrogen's to two parts in a trillion.

Einstein's theory of gravity shows none of these cracks. Its gravity pulls on energy, and swapping matter for antimatter leaves the energy the same, so the theory says antimatter falls at exactly 1 g.

So doesn't the triple mirror settle the drop? No. Apply it to an apple falling on the Earth and you get an anti-apple falling on an anti-Earth, made of antimatter too. On its own, it says nothing about an anti-apple on our Earth. That needed a separate rule, that everything falls alike, and in the end it needed the drop.

Three scenes

an apple, the Earthit falls

an anti-apple, an anti-Earthit falls the same way: the triple mirror says so

an anti-apple, our Earththe mirrors say nothing: it had to be measured

What about the film run backwards? A falling apple, played in reverse, flies up and slows down: it is still pulled down. Running time backwards never flips which way gravity pulls.

A popular picture treats an antiparticle as a particle following its path backwards in time. Picture it that way, and the electric push turns round, as if the charge had flipped; gravity's pull does not. So even that picture says antimatter falls down (this page's argument).

Flip the mirrors 👇

Whatever you switch on, the apple is pulled towards its Earth, and every mirror that turns the apple into antimatter turns the Earth into antimatter too. An anti-apple falling on our Earth isn't a scene that symmetry alone can make. It had to be measured.

Part threeWhat they are hunting now

Entry 11

If it isn't exactly 1 g

Suppose a future drop finds 0.99 g, with a tiny error bar. What would that mean?

One viewHow small a difference would be interesting? One theorist, Diego Blas, told CERN's magazine that things start to get interesting at about one part in ten million. That's one view.

Probably not that the triple mirror is broken: that rule is about an anti-Earth. The usual reading would be a new, very weak force that treats matter and antimatter oppositely, like the partners of gravity in chapter 6, or gravity not pulling equally on every kind of energy.

Such forces are already squeezed hard, by every matter-only test and by those pairs of neutron stars. And remember the best-known idea of such a force: antimatter falls a little faster. A real difference could go either way.

Guess first

The Higgs field gives the three antiquarks inside an anti-atom their mass. How much of the anti-atom's mass is that?

See the answer

About 1%. Read on for the rest.

What an anti-atom is made of matters too. Its antiproton is built from three antiquarks, tiny building blocks. The Higgs field (the field that gives basic particles their own mass) gives them theirs, but that is only about 1% of the anti-atom's mass. Even counting the short-lived quark pairs in there as well, quark masses make up only about a tenth. The rest is the energy of quarks and gluons (the particles that glue quarks together) moving and pulling (Nº 5).

So the drop mostly tests how gravity pulls on that energy. A few theorists argue that this leaves bare antiquarks untested.

And there's one thing the drop doesn't do: explain why the universe holds so much more matter than antimatter. That's a separate puzzle.

Entry 12

Twenty times sharper

The error bar, now and next

the 2023 error bar, about a fifth of g

the goal: 1% of g

Going from about a fifth of g to 1% means an error bar about twenty times smaller (this page's arithmetic).

ALPHA says its 2023 result was reproduced in a 2024 run, and that 1% or better should be possible with the same machine. The atoms are coming faster: in November 2025 the team reported more than 15,000 anti-atoms trapped in under seven hours in its other trap, thanks to a new way of chilling its positrons.

But counting is the smallest of the big pieces. In 2026 the team reported work in progress, in conference talks, on mapping the bottle's magnetic field and checking its detector, the two biggest pieces. As of 25 October 2026, as far as we can find, no new ALPHA-g number has been published.

Two other CERN teams, AEgIS and GBAR, are building different drops: one plans to send a beam of anti-atoms through fine gratings, the other plans to time a fall of 10 to 20 cm. Both aim at about 1%, and as of 25 October 2026, as far as we can find, neither has a result. For the 2030s there are sharper ideas still, such as fountains of anti-atoms that could reach about a part in a million.

All of it waits on the accelerators.

CERN's antimatter factory switched off for a long shutdown on 31 August 2026. CERN says its accelerators will come back in stages from 2028 onwards; as of 25 October 2026 it has given no date for the machines that slow the antiprotons down.

See how far the ladder still has to go 👇

The ladder makes the gap plain. ALPHA-g's error bar is as big as the weakening of gravity about 800 km above the ground. One part in ten million is like the change from the floor to the top of a 32 cm step stool (this page's arithmetic).

Entry 13

Always a long shot

So here are five things 0.75 g does not mean.

~~That antimatter falls slower.~~ 0.75, give or take 0.21, is consistent with 1 g.
~~That antimatter has negative mass.~~ Its mass and energy are positive, like matter's; only gravity's pull on it was open.
~~That "everything falls alike" is proved for antimatter.~~ It was tested to about a fifth. The indirect limits are far tighter, and each has an "if".
~~That the triple mirror forced the answer.~~ It speaks about an anti-Earth; our Earth had to be measured.
~~That the missing antimatter is explained.~~ The drop doesn't touch that puzzle.

Where things stand, as of 25 October 2026: ALPHA-g's 0.75 g is still the only direct observation of gravity acting on antimatter. The arguments that made "up" a long shot have now been joined by an observation. The arguments against a small difference are still arguments, each with its "if".

Let's write it all up. 📝

Lab notebook: antimatter's fall

Antihydrogen's fallALPHA-g, 20230.75 g

Give or takethis page's combination of ALPHA's two error bars± 0.21 g

Antigravity at full strengthALPHA-gruled out

Height of the bottlebetween the magnetic lids25.6 cm

Walls against gravitythis page's arithmetic1,600 : 1 (energies)

Anti-atoms counted per tryon average18–32

Morrison's machineif antimatter fell upenergy from nothing

Indirect limit, if one assumption holds2009 preprint1 in 10 million

Next goalALPHA, AEgIS, GBAR1%

Which way does antimatter fall?as far as anyone can measureDown

antimatter falls down ✦ almost nobody expected "up" ✦ exactly 1 g is the next question

In a quiet hall at CERN, a bottle of anti-atoms hung in the dark, held by nothing but magnetism.

Over twenty seconds, its two lids sank together.

One by one, the atoms slipped out and flashed on the walls.

More flashes below than above.

The oldest arguments in the book had said so. Now the counts did too.

Antimatter falls down. The question now is whether it falls at exactly 1 g.

Part IThe drop

Chapter 1Zero point seven five g

On 27 September 2023 the ALPHA collaboration at CERN reported in Nature the first direct observation of gravity acting on antimatter.¹ The team held atoms of antihydrogen, each an antiproton with a positron bound to it, in a vertical magnetic trap, lowered the magnetic barriers at its top and bottom together, slowly, and counted which way the atoms left. More left through the bottom. Repeating this at a series of deliberate magnetic tilts, and comparing the up/down split with simulations, gave antihydrogen's acceleration towards the Earth:

a_g = 0.75 ± 0.13 (statistical and systematic) ± 0.16 (simulation) gg = 9.81 m/s², the value ALPHA used; positive means towards the Earth. The second uncertainty comes from the simulation, chiefly from how well the trap's magnetic field is modelled away from its axis (Chapter 5).

Combining the two uncertainties in quadrature, as the square root of the sum of their squares:

√(0.13² + 0.16²) = 0.206 ≈ 0.21 g, a_g ≈ 0.75 ± 0.21 g = 7.4 ± 2.0 m/s²This page's arithmetic. Adding in quadrature assumes the two terms are independent.

So the headline number carries an error bar of about a fifth of g, the 20% that CERN quotes.² The gap between 0.75 and 1 is 0.25 g, about the size of the combined uncertainty: by this page's arithmetic, 1 g lies 1.2 standard deviations from the central value. That is what "consistent" means here.

What the result rules out, ALPHA states as probabilities. If gravity did not act on antihydrogen at all, the chance of a result at least this extreme would be 2.9 × 10⁻⁴: the probability of the data without gravity, not the probability that there is none. The probability that the data are consistent with repulsive gravity of 1 g is below 10⁻¹⁵, so small, the authors write, as to be "quantitatively meaningless".¹ Antigravity as strong as ordinary gravity is ruled out; subtler differences are not, and they are the subject of Part III.

Jeffrey Hangst, ALPHA's spokesperson, called it the first direct experiment to observe a gravitational effect on the motion of antimatter,^2,3 though ALPHA had set a loose limit in 2013 (Chapter 2).

Almost no physicist expected anything but "down", for reasons more than sixty years old. Philip Morrison (1958) imagined a perpetual-motion machine made of light;⁴ Leonard Schiff (1958–59) pointed to the virtual antimatter inside every atom;^5,6 Myron Good (1961), to neutral kaons (Chapters 6 to 9).⁷ In 1991 Michael Martin Nieto and Terry Goldman reviewed all of them and argued that a direct test was still needed.⁸

This page's thesis, then: the drop came out as nearly everyone expected, and it was worth doing because the expectations leaned on assumptions. What is left open is the size of the pull, not its direction.

Chapter 2Ninety years without a weighing

Antimatter entered physics through an equation. Paul Dirac wrote down a relativistic equation for the electron in 1928,⁹ and in 1931 proposed a particle with the electron's mass and the opposite charge, the "anti-electron"¹⁰ (Nº 11 meets the same equation inside a battery). Carl Anderson found it in cosmic rays and first reported it in Science on 9 September 1932, with the full paper in 1933.^11,12

The antiproton followed at Berkeley in 1955.¹³ In 1995 a team at CERN's LEAR ring made eleven candidate atoms of fast-moving antihydrogen, about two of which might have been background;¹⁴ the ATHENA experiment made cold antihydrogen in 2002;¹⁵ ALPHA trapped 38 atoms in 2010¹⁶ and reported laser cooling of trapped antihydrogen in 2021.¹⁷

One property was never in doubt: antimatter has ordinary, positive mass and energy. An electron and a positron at rest annihilate into two photons of 511 keV, carrying exactly the mass-energy of both (Nº 0), and in 2022 the BASE collaboration found the antiproton's charge-to-mass ratio equal to the proton's to 16 parts per trillion.¹⁸ Both measure the inertial mass: how hard antimatter resists being pushed. What no experiment had touched was the gravitational mass: how hard gravity pulls on antimatter. The data never left room for negative mass; what they left open was whether gravity pulls on antimatter differently.

Why not simply weigh a positron? Because the faintest stray electric field pushes on a charged particle far harder than gravity does. In 1967 Fred Witteborn and William Fairbank dropped electrons down a shielded tube and showed how hard that is.¹⁹ Their 1964 proposal had named positrons too,²⁰ and their 1968 Nature paper said a method for positrons was being devised,²¹ but the positron run never gave a result. As far as we can find, no charged antiparticle has ever been weighed directly.

Antihydrogen is neutral, but it must be held without touching anything, since contact with matter annihilates it. In 2013 ALPHA used the release of trapped atoms to set the first direct limit. Writing F for the ratio of antihydrogen's gravitational mass to its inertial mass (1 for ordinary gravity, −1 for antigravity), only −65 < F < 110 survived once systematic effects were included.²² Antimatter's weight could then still have been anything from 65 times normal pulling upwards to 110 times normal pulling down.

Chapter 3A magnetic bottle a quarter of a metre tall

For the 2023 drop the collaboration used ALPHA-g, an apparatus whose trap stands upright inside a solenoid, a long coil with a field of 1 tesla. Two short "mirror" coils, one at the bottom and one at the top, make field peaks of about 1.74 T on the axis, 25.6 cm apart; octupole windings, eight-pole coils round the sides, hold the atoms in. The nominal trapping volume is a cylinder 4.4 cm across and 25.6 cm tall, a bit shorter than a 30 cm ruler.¹

The atoms are held by their magnetism, not their charge, which is zero. A ground-state antihydrogen atom in the right spin state is a "low-field seeker": its energy rises in a stronger field, so it is pushed towards weak field, and the weak-field region between the mirror peaks is the bottle, the peaks its lids. The trap is about 0.5 K deep in temperature units.¹ The atoms are not dropped from rest: they bounce from lid to lid many times a second.

Against those walls gravity is tiny. Lifting a hydrogen atom, or an antihydrogen atom of the same inertial mass, through the height of the bottle costs

ΔU = m_HgL = 1.6735 × 10⁻²⁷ kg × 9.81 m/s² × 0.256 m = 4.20 × 10⁻²⁷ J ≈ 26 neV ≙ 0.30 mKm_H: the mass of a hydrogen atom (CODATA 2022 values²³); L: the distance between the mirror peaks. This page's arithmetic.

So the magnetic walls are about 1,600 times deeper than gravity's drop across the whole trap (this page's arithmetic; a ratio of energies, not of forces). Put another way, an atom of a given energy could coast upwards against gravity alone to a height

h = k_BT / (m_Hg): 0.5 K → 420 m; 100 mK → 84 m; 10 mK → 8.4 mGravity is worth 1.19 mK per metre of height for a hydrogen atom. This page's arithmetic.

The 420 m belongs to the most energetic atoms the trap can hold, not to a typical one. Even a 10 mK atom could climb about 8 m, and energies of that order can now be read from where and when atoms annihilate: a 2026 ALPHA paper reaches about 10% precision over mean energies of about 10 to 100 mK in the ALPHA trap.²⁴ A 20 mK atom crosses the bottle and back in about 28 ms (this page's arithmetic).

How can gravity show up at all? Every atom feels it, all the time, as the same small tilt: the bottom lid is lower than the top one by 0.30 mK. The trick is to lower both lids together, slowly. An atom escapes once its energy along the axis exceeds the combined magnetic and gravitational barrier, so it leaves over whichever lid it first clears as both sink,¹ and gravity gives the bottom lid a head start.

The lids come down steadily over 20 seconds (Chapter 4), about 25 mK a second. In one round trip of a 20 mK atom they sink about 0.7 mK, more than twice the 0.30 mK tilt, so the bottom lid's head start lasts less than half a round trip, and a slow atom heading upwards when the bottom lid sinks below its energy can still reach the top first (this page's arithmetic). Gravity sets the odds rather than deciding each atom: in ALPHA's simulation of ordinary atoms, about four in five leave by the bottom.¹ The measurement is statistical, like reading the bias of a loaded coin from many tosses, and ALPHA notes that colder atoms would make it more sensitive.

Chapter 4Opening both lids

Each trial began by gathering about 100 antihydrogen atoms in the bottle, made in about 50 rounds of mixing antiprotons with positrons over roughly four hours.¹ Then came two releases. First, the long octupole magnet (LOc) was ramped down in one second, letting go of the atoms moving fastest sideways, about half of the sample. They are not part of the gravity measurement: counting them tells the team how many atoms it had, and normalises the calibration of the detector's up and down regions (Chapter 5).

Second, the two mirror coils were ramped down together, linearly, over 20 seconds, releasing the rest: about 50 atoms per trial (this page's estimate: about 23 counted escapes per trial in Table 1, plus about 8 that came out after the ramp, divided by the detection efficiency of 0.65). The lids stopped not at zero but at a residual barrier of about 3.4 mK (a field about 5 × 10⁻³ T above the bottom of the well; this page's conversion). Almost nothing left in the first ten seconds, so ALPHA counted escapes from 10 to 20 seconds.

A radial time projection chamber, a gas-filled detector round the bottle that tracks the charged debris of each annihilation, located it on the wall to about 2 cm, with a detection efficiency of 0.65 ± 0.02. Averaged over a setting, a trial gave about 18 to 32 counted escapes; a setting, after six or seven trials, 123 to 195; all the settings together, 1,722. Another 621 events came out after the ramp; ALPHA left them out of the result because their systematic checks had not been done.¹

To trace a curve rather than a single split, ALPHA tilted the bottle on purpose. A small field difference between the mirror peaks raises one lid against the other, as gravity does. This deliberate tilt is the bias, quoted in units of g; a bias of 1 g is the field difference that matches gravity's own tilt:

ΔB_1g = m_HgL / μ_B = 4.53 × 10⁻⁴ T, a gradient of 1.77 × 10⁻³ T/mμ_B: the Bohr magneton, the atom's magnetic moment. The paper's values; this page's arithmetic gives 4.532 × 10⁻⁴ T and 1.770 × 10⁻³ T/m.

The biases were 0, ±0.5, ±1, ±1.5, ±2 and ±3 g (positive values push atoms down), over 75 trials, plus 12 calibration trials at ±10 g, where almost every atom leaves one way. ALPHA's simulation of ordinary atoms in an untilted trap sends about 80% out of the bottom, with the even split near −1 g, where the bias cancels gravity, as a simple one-dimensional estimate predicts.^1,2 Antigravity would put the even split near +1 g.

Bias	Counted up	Counted down	Share down	This page's rough correction
−10 g	142.9	0.7	0.005	calibration
−3 g	151.7	16.5	0.098	0.10
−2 g	128.7	33.5	0.207	0.20
−1.5 g	128.9	57.7	0.309	0.30
−1 g	69.7	62.5	0.473	0.47
−0.5 g	55.7	67.5	0.548	0.54
0	36.7	94.5	0.720	0.71
+0.5 g	36.7	124.5	0.772	0.77
+1 g	17.7	119.5	0.871	0.87
+1.5 g	13.9	180.7	0.929	0.93
+2 g	6.7	163.5	0.961	0.96
+3 g	7.7	147.5	0.950	0.95
+10 g	−0.1	185.7	≈ 1.00	calibration

Table 1 of Anderson et al. 2023: counted escapes from 10 to 20 s, background-subtracted (hence the fractions) but not efficiency-corrected. The last two columns are this page's arithmetic (Chapter 5).

Figure 1 runs one release in this page's simplified model, beside ALPHA's real tally: a single trial catches twenty-odd atoms with a jumpy split, and only whole settings trace the curve.

Figure 1🍾 Try it

Open both lids

Pick a tilt, then open the lids

Rings: counting noise only. The real error bar, ±0.21 g, is more than three times the counting term of 0.06 g.Rings show counting luck only; the real error bar is more than three times bigger.

±0.21 g (1σ)the error bar±0.41 g (2σ)twice the error barALPHA-g 2023, Table 1the real experimentyour simulated counts: counting noise onlyyour simulated atoms: counting luck only

Tilt (bias)Tilt

Time into the release

Calibration setting: ALPHA's counts onlyA calibration setting: only the real experiment's counts

Lids atLid height

This trial, counted up · downCaught at the top · bottom

Share leaving down, your simulated campaignShare out of the bottom

ALPHA-g 2023 hereThe real experiment here

Still in the bottle

One release in this page's illustrative model, beside ALPHA-g's published tally. The lids fall linearly from 0.5 K to a residual 3.4 mK in 20 s (energies in millikelvin, k_B × 10⁻³ K); about 50 atoms per trial (this page's estimate), escape energies log-uniform from 1 to 250 mK (capped at 250 mK so that, as ALPHA saw, almost nothing leaves before 10 s), detection 0.65. Gravity and the bias tilt both lids for every atom by 0.30 mK per g, the same gap at every moment of the ramp: the log gauge shows the lids' mean level, with each atom's tick placed where that level stands when the atom leaves, and the linear inset beside it shows the gap (true scale, or ×100 when magnified). In the model an atom escapes when the lower lid sinks below its energy; in the real trap it leaves over whichever lid it first clears as both sink. Why a slow atom can still leave by the top: Chapter 3. Which lid an atom leaves by is drawn from this page's toy curve P_dn = f + (1 − 2f)/(1 + e^−(b+a)/w), f = 0.05 (a floor), w = 0.621 g (set so that ordinary atoms give ALPHA's simulated 0.80 at zero bias; Chapter 5), with a the pull you pick: this page's toy, not ALPHA's simulation, which uses derived biases. Each simulated setting gets its own atoms; rings show counting noise only, with as many counted escapes per setting as Table 1 (123–195). At the ±10 g calibration settings only ALPHA's counts are shown. Points: ALPHA-g 2023, Table 1, background-subtracted counts, share down (Anderson et al. 2023); the alternative view applies this page's rough up/down efficiency correction from the ±10 g calibration trials, which moves the points slightly down; ALPHA's own correction moves them slightly up. Bands: darker a = 0.54–0.96 g (±0.21 g, 1σ); lighter 0.34–1.16 g (±0.41 g, twice 0.206 g; this page's arithmetic). The toy curve is drawn from −3 to +3 g only.Open the lids and nothing happens for ten seconds; then atoms slip out: with no tilt and ordinary gravity, mostly through the bottom. One try catches only twenty or so, so the split jumps about: run a whole setting, with as many atoms caught as the real team caught there, or the whole set. The gauge's numbers are energies, written as temperatures in thousandths of a degree (millikelvin, mK). Gravity tilts the lids by the same tiny amount all the way down; the small bar beside the gauge shows it. Which end an atom reaches first as the lids sink is partly luck; gravity only tips the odds. The darker band is the measurement's error bar; the lighter band is twice that, and ordinary gravity sits inside it. In the chart of how many left by the bottom, the dots are the real experiment's counts and the rings are your simulated ones.

ALPHA's raw share at zero bias, 0.72, sits below the 0.80 that its simulation gives ordinary atoms. That is not, by itself, a sign of weaker gravity: the counts are small, the calibration of the detector's up and down regions is itself uncertain, and the real biases differ slightly from the nominal ones (Chapter 5).

Chapter 5Where the error bar comes from

ALPHA's analysis fits full simulations of escaping atoms to the up/down shares at every setting. The simulations use "derived" biases: the tilts worked out from measurements and models of the actual fields, rather than the nominal ones.¹ This page cannot repeat that analysis, but two rough steps show its shape.

First, an efficiency correction. At ±10 g almost every atom leaves one way, so comparing those counts with the LOc count gives each region's relative efficiency, about 0.846 counted escapes per LOc event for up and 0.872 for down (not the 0.65 detection efficiency of Chapter 4). Correcting Table 1 with them lowers the share down at zero bias slightly, from 0.72 to about 0.71, and puts the even split near −0.8 g (this page's rough correction, the last column of Table 1, not ALPHA's analysis). ALPHA's own corrected shares, plotted in its Fig. 5 from a fit that also uses the LOc counts at every setting, move the other way, up to about 0.74 at zero bias (this page's estimate, from a fit of the same kind to Table 1): the two estimates of the efficiency ratio differ by less than its counting error (below).

Second, a toy curve for the share leaving by the bottom: an S-shaped step centred where the bias cancels antimatter's pull, with a small floor f for atoms that leave by the "wrong" lid even at large tilts (Table 1 at ±3 g). Figure 1 uses it:

P_dn(b, a) = f + (1 − 2f) / [1 + e^{−(b + a)/w}], f = 0.05, w = 1/ln 5 = 0.621 gb: the bias; a: antimatter's acceleration, both in units of g; f: that floor. The width is chosen so that ordinary atoms (a = 1) give 0.80 at zero bias and 0.50 at −1 g, as ALPHA's simulation does; only a is then fitted, to this page's corrected shares. This page's toy, not ALPHA's analysis.

Fitted to the corrected shares, the toy's best value of a is about 0.75, ALPHA's central value, but the agreement is partly luck, not a check: corrected ALPHA's way, the shares move slightly up, and the toy's value with them. What matters is the uncertainty, which ALPHA's Table 3 breaks down in two blocks.¹

Source of uncertainty on a_g	1σ, in g	Block
Finite data size (counting)	0.06	statistical and systematic
Calibration of the detector efficiencies in the up and down regions	0.12	statistical and systematic
Other minor sources	0.01	statistical and systematic
Block total, in quadrature	0.135 → 0.13
Modelling of the magnetic fields, on and off the axis	0.16	simulation
Initial energy distribution of the atoms	0.03	simulation
Block total, in quadrature	0.163 → 0.16

The quadrature sums are this page's arithmetic. Other terms make the tilts themselves uncertain; a shift of the curve along the bias axis moves the fitted acceleration one for one (the width of the electron-cyclotron-resonance spectrum used to measure the field, 0.07 g; field modelling, 0.05 g; and smaller ones), and the octupole's field could shift the curve by up to ±0.26 g along that axis. The paper folds all of these into its quoted uncertainties; the octupole's shift enters the simulation band as a flat spread from −0.26 to +0.26 g, whose standard deviation is 0.26/√3 ≈ 0.15 g (this page's arithmetic).

Counting is the smallest of the three large terms in the table. The efficiencies come from the ±10 g runs, which had counting errors of their own: each efficiency rests on two counts of 143 to 213 events, so the up/down efficiency ratio is uncertain by about 15% from counting alone, enough to move this page's toy fit by about 0.13 g (this page's arithmetic). Much of the 0.12 g may therefore be counting in the calibration runs; the paper does not say how much.

So an error bar built from counting alone would mislead. With the counting term only, 0.06 g, the 0.25 g between 0.75 and 1 would look like a deviation; with the full ±0.21 g, more than three times larger, it does not. Figure 2 lets you shrink each piece.

Figure 2📏 Try it

Where the error bar comes from

Shrink the pieces

More gravity-run atoms (hypothetical; calibration runs unchanged)

Better field map

Better up/down calibration

Better-known starting energies

Total error barGive or take

Share of g

Biggest piece

Still to go to 1%

A scale this uncertain would say

ALPHA-g's published uncertainty terms on the acceleration (Anderson et al. 2023, Table 3). Statistical and systematic, ±0.13 g: finite data 0.06 g, up/down detector-efficiency calibration 0.12 g, other minor sources 0.01 g. Simulation, ±0.16 g: magnetic-field modelling 0.16 g, initial energy distribution 0.03 g. The four largest are combined here in quadrature (this page's arithmetic); the 0.01 g term is not modelled, and a real 1% measurement must shrink it too. The sliders divide each term by an improvement factor. The atoms slider multiplies the gravity-run data by m and shrinks the finite-data term as 1/√m; the calibration runs stay as they were. ALPHA calibrates the efficiencies from its ±10 g runs, and the paper does not say how much of the 0.12 g is counting in them. Bias-axis terms (0.07, 0.05 g) and the octupole's possible ±0.26 g shift are not shown separately: ALPHA carries them into its published terms. The 2026 conference talks (work in progress) target the two largest sliders: mapping the trap's field by electron cyclotron resonance (CAP Congress, June 2026) and position- and rate-dependent detector efficiencies (WNPPC, February 2026); the beryllium cooling reported in 2025 (more than 15,000 atoms in under seven hours in ALPHA's spectroscopy trap, and an initial gain of more than twenty times in ALPHA-g) bears on the atoms slider. Improvement factors are hypothetical; ALPHA expects 1% or better in the existing machine (Caravita et al. 2025).The measurement's error bar is made of five pieces; four are on sliders (the smallest, a hundredth of g, isn't). Shrink each one and watch the total. More atoms in the main runs only shrinks the counting piece, the smallest of the three big ones. The scale row shows what an error this size would mean on a bathroom scale.

Two lessons follow. More gravity-run atoms alone barely help: a thousand times the data takes the counting term to 0.0019 g but the total only to ±0.20 g, and even if the calibration runs grew too, the total could not fall below about ±0.16 g, the floor set by the simulation terms. Reaching 1% means cutting every piece, Table 3's 0.01 g of other sources included: that term is itself 1% of g, so even with the other four cut to 0.8% of g between them, as in Figure 2's 1% preset, it would hold the total near 1.3% (this page's arithmetic).

What 0.75 g does not mean, then, is "antimatter falls three quarters as fast". It is the central value of a measurement uncertain by about a fifth of g, and 1 g lies inside twice that uncertainty, the band from 0.34 to 1.16 g.

Part IIWhy "up" was always a long shot

Chapter 6A perpetual-motion machine made of light

In 1958 Philip Morrison gave the oldest and simplest argument against antigravity.⁴ Suppose antimatter fell up. Lift a particle and its antiparticle together up a tower: the matter half resists, the antimatter half pulls upwards just as hard, so the lift costs nothing. At the top, let them annihilate and send the light down the tower. Light gains energy as it falls, so at the bottom it can re-make the pair with energy to spare: a perpetual-motion machine.

The argument bites because light really does fall. Robert Pound and Glen Rebka measured photons gaining energy down a 22.5 m tower in 1960, about 2.5 parts in 10¹⁵, to about 10%;²⁵ in 2022 a JILA team measured the same shift across a cloud of strontium atoms about a millimetre tall.²⁶ Near the ground the fractional gain is gh/c², 1.09 × 10⁻¹⁶ per metre (Nº 9 is about this shift). And the photon is its own antiparticle, so light cannot fall one way for matter and another way for antimatter.

With m the mass of one member of the pair and a antimatter's downward acceleration, the ledger fits on one line:

lift: m(g + a)h; light's gain: 2mc² × gh/c² = 2mgh; net per cycle = mh(g − a)Annihilation at the top gives 2mc² of light; re-making the pair at the bottom costs 2mc². For a = −g (antigravity) the net is 2mgh; for a = g it is zero. This page's arithmetic.

With antigravity, a hydrogen–antihydrogen pair lifted one metre leaves 3.3 × 10⁻²⁶ J (0.20 µeV) over every cycle; a kilogram of pairs leaves 9.8 J per metre from nothing, the energy needed to lift a 1 kg bag of sugar by a metre (this page's arithmetic; for hydrogen, turning all the energy into light is an idealisation).

Our argument: any difference, not only "up". The net mh(g − a) vanishes only at exactly a = g. Below g the cycle leaves energy over; above it, the cycle run backwards (lower the pair, annihilate at the bottom, send the light up) gains mh(a − g). So within this logic any difference from 1 g, either way, makes a perpetual-motion machine unless a new field stores or supplies the energy. The arithmetic is this page's; the idea is not. Fischler, Lykken and Roberts state Morrison's argument for an antiparticle merely "deficient" in its attraction,²⁷ and in 1975 Kenneth Nordtvedt gave the general form: energy conservation ties any difference in free-fall rates to a matching anomaly in the gravitational redshift.²⁸

If antimatter fell at exactly 0.75 g with no new field, each one-metre cycle would leave 4.1 × 10⁻²⁷ J (26 neV) per hydrogen pair. That is what energy conservation would forbid, not what ALPHA measured: 0.75 ± 0.21 g is consistent with 1 g, where the books balance. Figure 3 runs the cycle for any value you choose.

Figure 3💡 Try it

Morrison's ledger

Set antimatter's fall, run the cycle

Antimatter's fall

−1 g, up01 g

±0.21 g (1σ)the error bar±0.41 g (2σ)twice the error bar

Lifting the pair costsLifting costs

The light gains on the way downLight picks up

Left over each cycleFree energy each cycle

A free lift of the whole pair byLike lifting the pair for free by

Morrison's (1958) cycle as an energy ledger: lift a particle–antiparticle pair through h, annihilate it and send all of its energy down as light (an idealisation for the hydrogen pair; an electron–positron pair annihilates into light directly), let the light fall (gaining the fraction gh/c², 1.09 × 10⁻¹⁶ per metre, Nº 9) and re-make the pair. Lift m(g + a)h, light's gain 2mgh, net mh(g − a), with m one member's mass and a antimatter's downward acceleration. The net vanishes only at a = g. That any a ≠ g leaves energy over is already in Fischler, Lykken and Roberts (2008) and, in general form, Nordtvedt (1975); the ledger arithmetic and its numbers are this page's. It assumes no new field stores the energy: Nieto and Goldman (1991) and Scherk (1979) show how one could. Shading: ALPHA-g's 0.75 ± 0.21 g and ±0.41 g, twice 0.206 g (0.34–1.16 g).Lift a matter–antimatter pair, turn all of its energy into light at the top, let the light fall and turn it back into a pair. Slide how antimatter falls. Only at exactly ordinary gravity does the energy come out even; anything else makes energy from nothing, unless an unknown force is hiding it.

Morrison did not prove antigravity impossible: he showed that it would break energy conservation in Einstein's theory, and so where a loophole would have to be.

Chapter 7Where the energy could hide

Nieto and Goldman's 1991 review took that loophole seriously. As summarised by Fischler, Lykken and Roberts, they concluded that Morrison's argument goes through in pure tensor gravity, Einstein's theory with nothing added, but can fail if gravity has partners.^8,27 Two partners of the graviton, the hypothetical carrier of gravity, could make antimatter fall differently without breaking energy conservation: a "gravivector" (spin 1) and a "graviscalar" (spin 0), if both coupled not to mass-energy but to baryon number (the count of protons and neutrons, with antiprotons and antineutrons counting minus one) or to lepton number (the same count for electrons, muons, taus and neutrinos).

Those counts act as charges, and they change sign for antimatter: that is the point. A vector force repels like charges and attracts opposite ones, as in electromagnetism; a scalar force attracts either way. For matter falling on the matter Earth the two could nearly cancel, hidden from ordinary experiments; for antimatter the vector part changes sign and the two add. The new fields would then balance Morrison's books.

The best-known model of this kind points the other way from "up". In 1979 Joël Scherk argued that the gravivector partners of the graviton in supergravity theories, extensions of Einstein's theory, would make antimatter fall slightly faster than matter, with no dissipation or instability (Chapter 12).^29,30

A second escape keeps pure tensor gravity and calls the machine too feeble to matter. Gabriel Chardin and Jean-Marcel Rax argued in 1992 that the instability of the vacuum that antigravity would imply is of the kind and size of Hawking evaporation, the faint glow Stephen Hawking predicted from black holes, unnoticeable in a weak field like the Earth's;³¹ Chardin and Giovanni Manfredi restated it in 2018.³⁰ It is a minority position; Chardin is also an author of the Dirac–Milne cosmology of Chapter 10. So Morrison's argument rules out antigravity unless new forces carry the energy or, on this minority view, the machine is too feeble to notice.

Chapter 8The antimatter already inside every atom

Leonard Schiff's argument, published in 1958 and expanded in 1959, starts from quantum electrodynamics.^5,6 The electric field of a nucleus is dressed with short-lived virtual electron–positron pairs, the effect called vacuum polarisation, which alter its electric energy, and so its mass, by amounts that differ from element to element. They are not real positrons sitting in the atom, but if their share of the energy fell up, different materials would fall at different rates. Eötvös-type experiments, which compare how different materials fall, saw no such difference, and Schiff concluded that the positron's gravitational mass is positive.

Nieto and Goldman, as Chardin and Manfredi report, noted that the calculation did not handle correctly the infinities that quantum field theory has to subtract (renormalisation).^8,30 A minority objection comes from Chardin and Manfredi: building on Richard Price's 1993 analysis of a positive and a negative mass bound together,³² they argue that in such a model an exactly balanced pair floats and any unbalanced pair falls normally, so a violation would be all or nothing.³⁰ The argument treats antimatter as negative mass, which is not the mainstream view and sits awkwardly with its measured positive inertia (Chapter 2).

In 2009 Daniele Alves, Martin Jankowiak and Prashant Saraswat made Schiff's idea quantitative, in a preprint that, as far as we can find, never appeared in a journal.³³ Their stated assumption is that any matter–antimatter difference in gravity shows up equally in the virtual pairs around nuclei and in the "sea" of virtual antiquarks inside protons and neutrons. Under it:

the Lamb shift, the small quantum correction to atomic energy levels to which virtual pairs contribute, limits a difference to about 10⁻²;
the vacuum-polarisation part of a nucleus's electrostatic self-energy (the energy of its protons pushing on one another) is about 10⁻³ of that energy, or about 10⁻⁶ of the atom's mass, and it differs by about 10⁻⁶ between beryllium and titanium. The Eöt-Wash torsion balances, twisting pendulums that see no difference between how those metals fall at about 10⁻¹³, therefore limit a difference to about 10⁻⁷;
the sea of virtual antiquarks inside protons and neutrons, about a tenth of their energy, would give about 10⁻⁹.

The torsion-balance input is the Eöt-Wash group's beryllium–titanium result, η = (0.3 ± 1.8) × 10⁻¹³,³⁴ part of a series reviewed in 2012.³⁵ Since then the MICROSCOPE satellite has found titanium and platinum falling alike to within a few parts in 10¹⁵, η = (−1.5 ± 2.3 ± 1.5) × 10⁻¹⁵³⁶ (Nº 1). Rerun with those two metals (nuclear electrostatic energies of about 91.5 and 745.7 MeV, from E ≈ 0.72 Z(Z − 1)/A^1/3 MeV), the self-energy estimate gives about 1–3 × 10⁻⁹ under the same assumption: this page's arithmetic, not a published number. The published limit is Alves and colleagues' 10⁻⁷, from Eöt-Wash.

Every one of these numbers carries the same "if". Experiments on ordinary matter have not shown that antimatter falls normally to one part in ten million. They have shown that any difference appearing in the virtual antimatter of nuclei, in the same proportion as in real antimatter, must be smaller than that.

Chapter 9Kaons, clocks and a supernova

Kaons. In 1961 Myron Good turned to neutral kaons, particles made of a quark and an antiquark.⁷ A K⁰ and its antiparticle turn into each other, and the particles that travel are mixtures of the two, the long-lived K_L and the short-lived K_S, whose masses differ by (3.484 ± 0.006) × 10⁻¹² MeV.³⁷ Under antigravity the kaon's energy would shift by m_KΦ one way and the antikaon's the other, with Φ the gravitational potential, the energy per unit mass in a gravity well; the difference, 2m_KΦ, would dwarf that splitting. With the Earth's potential, Φ/c² = 6.96 × 10⁻¹⁰:

m_KΦ ≈ 497.6 MeV × 6.96 × 10⁻¹⁰ ≈ 0.35 eV ≈ 1 × 10⁵ Δm_KThis page's arithmetic from PDG 2026 values; the difference is twice this, about 2 × 10⁵ Δm_K. With the Sun's potential at the Earth the energy is 4.9 eV (1.4 × 10⁶ Δm_K); with the Galaxy's, about 290 eV (8 × 10⁷ Δm_K).

Long-lived kaons would then regenerate short-lived ones, wrecking the observed pattern of kaon decays. The weak spot is the choice of potential: Earth, Sun or Galaxy give answers a thousandfold apart, and an absolute potential is not something experiments measure. In 1992 Chardin and Rax, and separately Goldman, Nieto and Vernon Sandberg, restated the argument without an absolute potential.^31,38 Goldman and colleagues found the effects generally very small, though measurable in principle; Chardin and Rax argued that antigravity would then give an effect of about the size of the observed CP violation in kaons (Chapter 11).^30,31 In its original form the kaon argument points down; without an absolute potential it is much weaker.

Clocks. A trapped antiproton circles at a frequency set by its charge-to-mass ratio, so it can be compared as a clock with a proton, or, as BASE does, with a negative hydrogen ion standing in for one, a line begun by Richard Hughes and Michael Holzscheiter in 1991.³⁹ BASE's 2022 measurement ran over a year and a half, as the Earth's distance from the Sun, and so the Sun's potential, changed. It fixed the ratio to 16 parts per trillion; from the change alone, the two clocks respond to the Sun's potential alike to within 0.030, about 3% (the "differential" limit). Tighter limits, below 1.8 × 10⁻⁷ in 2022¹⁸ and 8.7 × 10⁻⁷ in 2015,⁴⁰ rest on an assumed absolute background potential, which is disputed. BASE timed the antiproton; it did not weigh it. The link to free fall is Nordtvedt's (Chapter 6).²⁸

A supernova. The burst of neutrinos from Supernova 1987A, seen by the Kamiokande-II and IMB detectors, was mostly electron antineutrinos. A mass delays any signal that passes near it, the Shapiro delay; for this journey through the Milky Way, Michael Longo estimated some 4.8 months. The antineutrinos and the light were delayed by the same amount to within about 0.2–0.5%^41,42 (Nº 7 uses the same delay). Minazzoli and colleagues showed, though, that such an absolute delay is not something general relativity, Einstein's theory of gravity, defines on its own, so the limit is best read as an illustration.⁴³

All three clues point to "down", and each carries its own assumption: an absolute potential, an assumed background, an absolute Shapiro delay.

Chapter 10Why test it anyway

The ALPHA paper itself states where opinion stood: "The theoretical consensus is that any laboratory mass must be attracted by the Earth, although some authors have considered the cosmological consequences if antimatter should be repelled by matter."¹ That consensus was not the end of the matter, for two kinds of reason.

The first reason was method. Nieto and Goldman wrote their review to motivate a direct experiment.⁸ In theories with nearly cancelling forces (Chapter 7), they wrote, as quoted by Fischler, Lykken and Roberts, that "the more precisely anomalous gravitational effects are ruled out in earth-based matter-matter experiments, the more unrestricted is the possibility that there can be a significant anomalous gravitational acceleration of antimatter."²⁷

In 2008 Mark Fischler, Joseph Lykken and Thomas Roberts were asked by a Fermilab committee whether a 1% difference was already ruled out. Their note, which was not refereed, found that indirect reasoning "rests on models and assumptions which are in question for other reasons and are thus worth testing", and that "There is no compelling evidence or theoretical reason to rule out such a difference at the 1% level."²⁷ They added: "Indirect evidence does suggest, however, that the value is one to a much greater precision than the 1% level, 10⁻⁴ or perhaps as small as 10⁻⁸." Nothing directly excluded 1%; they did not say 1% was likely.

Alves and colleagues answered the cancellation point: "Even taking such cancellation for granted", the radiative damping of binary pulsars (the slow loss of orbital energy by pairs of neutron stars) "constrains these forces to be significantly weaker than gravity", to about 10⁻⁴; and imperfect cancellation tightens that further, to about 10⁻⁸ because the two forces' charges cannot match exactly and to about 10⁻⁷ because the forces depend differently on speed.³³

The second reason was that a few theorists took repulsion seriously, mostly for cosmology. Massimo Villata argued in 2011 that CPT (Chapter 11) applied to general relativity makes matter and antimatter repel;⁴⁴ Marcoen Cabbolet criticised the derivation.⁴⁵ Aurélien Benoit-Lévy and Gabriel Chardin proposed the Dirac–Milne universe, a cosmology built on matter and antimatter repelling each other;⁴⁶ Dragan Hajdukovic proposed virtual "gravitational dipoles" in the quantum vacuum, as a possible solution to the dark-energy problem.⁴⁷

ALPHA-g ruled out the simplest versions, and some have since revised their models. Villata's 2024 paper keeps cosmic repulsion in two forms that leave antihydrogen attracted.⁴⁸ Hajdukovic's refereed 2025 review reads the result as indicating "that atoms and anti-atoms have different gravitational charges", quotes "only 0.75 of that of ordinary matter" without its uncertainty, and adds "If confirmed by more precise measurements".⁴⁹ He treats it as a hint awaiting better data; the result itself is consistent with 1 g. His 2024 preprint, which argues that a slightly slower fall remains possible, had no journal reference as of 25 October 2026.⁵⁰ In a 2025 preprint, condensed for the Moriond Gravitation proceedings, Piacentino and Palladino argue that a composite atom does not test the bare antiquarks it is built from (Chapter 12), and that the result "may actually be consistent with a repulsive gravitational component of about 15%".⁵¹ These are minority readings beside ALPHA's own analysis.

That is the sense of this page's title. "Up" was always a long shot: almost nobody expected it, and the arguments against it were strong. It was never ruled out, because each argument rested on an assumption. The drop settles up versus down at the strength of ordinary gravity. It leaves small differences open.

Part IIIExactly 1 g?

Chapter 11Three mirrors and an apple

Physicists describe matter and antimatter with three mirrors. C, charge conjugation, swaps every particle for its antiparticle; P, parity, reflects space; T, time reversal, runs the film backwards. The CPT theorem, due to Gerhart Lüders in 1954 and Wolfgang Pauli in 1955, says that any local quantum field theory that respects special relativity is unchanged by all three together. It forces a particle and its antiparticle to have equal masses and lifetimes, and opposite charges and magnetic moments.

The weak force cracks each mirror on its own, and the pair CP too: P in the beta decay of cobalt-60 (Chien-Shiung Wu and colleagues, 1957);⁵² C because the weak interaction acts only on left-handed neutrinos and right-handed antineutrinos; CP in neutral kaons (Christenson, Cronin, Fitch and Turlay, 1964);⁵³ T in a 1998 CPLEAR kaon asymmetry that is also CP-violating, whose reading was debated,⁵⁴ and in B mesons at BaBar in 2012, by exchanging initial and final states.⁵⁵ The three together have never been seen to crack.

Antimatter tests of CPT include antihydrogen's 1S–2S line, the ultraviolet step from its lowest level to the next, matching hydrogen's to 2 parts in 10¹²;⁵⁶ its hyperfine splitting, the tiny energy gap set by the positron's and antiproton's magnetism, to 4 parts per million (ALPHA, May 2026);⁵⁷ the antiproton's magnetic moment, to 1.5 parts per billion;⁵⁸ its mass relative to the electron's, to 8 parts in 10¹⁰;⁵⁹ and the charge-to-mass ratio of Chapter 2.¹⁸ These are CPT tests, not gravity tests.

Einstein's gravity shows none of these cracks: it is unchanged by P and by T, and couples to energy and momentum, which C leaves unchanged, so general relativity predicts that antimatter falls at exactly g. Does CPT then guarantee "down"? No. Applied to an apple falling on the Earth, it gives an anti-apple falling on an anti-Earth, and says nothing on its own about an anti-apple on our Earth. That needs the weak equivalence principle, the rule that everything falls alike, which is related to CPT but not the same.^8,60

One popular picture treats an antiparticle as a particle travelling backwards in time, which can suggest that it should fall up. It should not. Film a falling apple and play it backwards: the apple rises and slows, still accelerating downwards. Acceleration does not change sign when the film runs backwards, for any motion. The backwards-in-time picture reverses only the particle's own clock, its proper time τ, in fields that stay as they were, and the relativistic equations show what that does:

d²x^μ/dτ² + Γ^μ_αβ (dx^α/dτ)(dx^β/dτ) = (q/m) F^μ_ν dx^ν/dτLeft: free fall in curved spacetime (the geodesic equation), quadratic in the four-velocity dx/dτ, so unchanged by τ → −τ. Right: the electromagnetic force, linear in the four-velocity, so it flips sign under τ → −τ, exactly as it would under q → −q.

In relativity the electromagnetic force is proportional to the particle's velocity through spacetime, including its direction through time, so reversing the particle's clock flips it, as flipping the charge would. Gravity's term contains that velocity twice, so the flip cancels. (The electric force qE itself does not depend on velocity; the flip comes through the direction in time.) So the backwards-in-time picture predicts that antimatter falls down (this page's argument). Villata's opposite conclusion (Chapter 10) is a minority view.^44,45

Figure 4 puts the three mirrors on switches; watch which way the apple is pulled.

Figure 4🍏 Try it

Three mirrors and an apple

Switch the mirrors on

Mirrors on

The mirrored sceneWhat you'd see

Allowed by the laws?Does nature allow it?

Gravity's pullWhich way it's pulled

Try the C switch on the apple.

C exchanges particles and antiparticles, P reflects space, T reverses the direction of time. The CPT theorem (Lüders 1954, Pauli 1955) makes the combination exact in any local Lorentz-invariant quantum field theory; each of C, P, CP and T is violated by the weak interaction (Wu et al. 1957; the V−A form of the weak interaction; Christenson et al. 1964; CPLEAR 1998, debated, and BaBar 2012). Einstein's gravity is unchanged by P and by T and couples to energy and momentum, which C leaves unchanged. CPT relates an apple falling on the Earth to an anti-apple falling on an anti-Earth; an anti-apple on our Earth is the weak equivalence principle, a separate assumption (Nieto and Goldman 1991; Kostelecký and Tasson 2011). Icons are illustrative.Three mirrors: swap matter for antimatter, flip left and right, run the film backwards. Switch them on one at a time, or all together, and see what nature allows. Notice that the antimatter switch always turns the Earth into antimatter too.

Whatever you switch on, the apple accelerates towards its Earth, and every mirror that turns the apple into antimatter turns the Earth into antimatter too. An anti-apple falling on our Earth is not a scene that symmetry alone can produce. It had to be measured.

Chapter 12What a difference would mean

Suppose a future drop finds 0.99 g with a small error bar. That would most likely not mean that CPT is broken, since CPT concerns an anti-Earth. The standard reading is a new, very weak, long-range force coupled to something that changes sign for antimatter, such as baryon or lepton number (Chapter 7), or gravity not coupling universally to all forms of energy.⁸ The Standard-Model Extension of Alan Kostelecký and colleagues gives such couplings a common language.⁶⁰

Such forces are already squeezed hard: by every matter-only test of free fall (Chapter 8) and, for nearly cancelling ones, by binary pulsars (Chapter 10).³³ And in the classic concrete model, Scherk's supergravity, antimatter falls slightly faster than matter (Chapter 7).²⁹ A real deviation could go either way.

What antihydrogen is made of matters too. Only about 1% of its mass is the Higgs-given mass of the three "valence" antiquarks that make up its antiproton, and about 0.05% its positron; counting the sea of virtual quark–antiquark pairs, quark masses make up about a tenth (9.3% to 11.1% on the estimates Nº 5 compares). The rest is the energy of quarks and gluons moving and pulling, as in hydrogen (this page's framing), so ALPHA-g mostly tests how gravity pulls on that energy. Piacentino and Palladino argue that this leaves bare antiquarks untested;⁵¹ a muonium test would involve no quarks at all (Chapter 13).

How small would an interesting difference be? One theorist, Diego Blas, told the CERN Courier in 2023 that things "start to get interesting when the precision reaches about one part in 10 million".³ That is one view; it is the level the indirect arguments of Chapter 8 reach under their assumptions, and ten times beyond even the 2030s fountain goals (Figure 5).

The measurement also says nothing about why the universe holds more matter than antimatter; the cosmologies built on repulsion are a separate question (Chapter 10).

Chapter 13Twenty times sharper

Going from ±0.21 g to 1% means an error bar about twenty times smaller (this page's arithmetic). ALPHA says its 2023 result was reproduced in a 2024 run and that a precision of 1% or better should be possible with the existing machine. From CERN's current long shutdown onwards, it plans upgrades to extract antihydrogen from the trap and use atom interferometry, which splits and recombines each atom's quantum wave so that gravity shifts the pattern. These are the collaboration's expectations, in the antimatter community's 2025 input to the European Strategy for Particle Physics, which was not refereed.⁶¹

The atoms are coming faster: in November 2025 ALPHA reported more than 15,000 antihydrogen atoms trapped in under seven hours in its spectroscopy trap, an eightfold increase in the trapping rate, by cooling its positrons with laser-cooled beryllium ions; added to ALPHA-g in 2024, the same cooling first raised its accumulation rate more than twentyfold, and CERN says the 2025 gravity runs used the new numbers.^62,63 But counting is the smallest of the three big pieces, and the 2026 work aims at the two biggest. Two talks report on mapping the trap's fields by electron cyclotron resonance (A. Powell, CAP Congress, June)⁶⁴ and on detector efficiencies that vary with position and rate (D. Duque, WNPPC, February).⁶⁵ Powell's thesis covers the fields;⁶⁶ the annihilation-pattern method of Chapter 3 can also probe the field away from the axis, its authors say;²⁴ and a 2025 report gives the detectors' status.⁶⁷ None of it is a result yet: as of 25 October 2026 no new ALPHA-g value appears among the 110 papers citing the 2023 result on the INSPIRE database.⁶⁸

Two other CERN experiments are building different drops. AEgIS plans to send a pulsed, horizontal beam of anti-atoms through a moiré deflectometer, fine gratings whose shadow pattern shifts as the atoms fall, aiming at about 1%.⁶⁹ It made pulsed antihydrogen in 2018 (published in 2021)⁷⁰ and reported laser cooling of positronium (an electron bound to a positron) in 2024.⁷¹ In March 2025 it judged a first measurement feasible before the shutdown;⁶¹ as of 25 October 2026 no AEgIS gravity result appears among the papers citing ALPHA-g, and its 2025 conference paper describes plans and development status only.^72,68

GBAR plans to make the positive antihydrogen ion, an antiproton with two positrons, cool it with laser-cooled beryllium ions, strip one positron with a laser and time a 10–20 cm fall, aiming at 1%.^61,73 It then hopes to reach about 10⁻⁵ with the quantum states of atoms bouncing just above a surface⁶¹ (a 2022 talk said 10⁻⁶⁷⁴). In its 2022 run it saw antihydrogen made in flight, some 20 atoms, a 3–4σ excess over background.^73,75 In 2025 the ion had yet to be produced, and as of 25 October 2026 we can find no report that it has been; a September 2026 preprint gives the first measurement of how readily its antiproton beams make antihydrogen.⁷⁶

For the 2030s the goals are sharper: fountain-type atom interferometry, which ALPHA's paper says could reach about 10⁻⁶, a part per million,¹ and the Canadian HAICU project, aiming at an anti-atom fountain.⁶⁵ Muonium, an electron bound to an antimuon (a heavier, short-lived cousin of the positron), offers a different test: a 2025 preprint describes an intense beam for a test at the 1% level with no quarks at all.⁷⁷

All of it waits on the accelerators. CERN's Long Shutdown 3 began when the LHC stopped in June 2026; the injectors and the Antimatter Factory, whose AD and ELENA decelerators slow the antiprotons down for the experiments, switched off on 31 August 2026.^78,79 CERN says the complex will return in stages "from 2028 onwards"; as of 25 October 2026 it gives no date for the antiproton decelerators.⁷⁹

Figure 5 puts these goals, the direct results and the indirect limits on one ladder of precision.

Figure 5🔬 Try it

How small a difference could hide?

Slide a hidden difference

ALPHA 2013: a direct bound, −65 < F < 110 (about 10²).ALPHA 2013: the first, loose direct limit, about 100 times g.
ALPHA-g 2023: direct, ±0.21 g (1σ; this page's combination).ALPHA-g 2023: the real drop's error bar, 0.21 g.
BASE, clock test, differential: 0.030 (a redshift test, linked to free fall through energy conservation).BASE's clock test, from the change alone: 0.030, about 3% (a clock test, linked to falling through energy conservation).
The 1% goals of ALPHA-g, AEgIS, GBAR's free fall and a muonium test (as of October 2026); Alves et al.'s Lamb-shift limit, 10⁻², if their assumption holds.The 1% goals of ALPHA-g, AEgIS, GBAR's drop and a muonium test (as of October 2026); the 2009 preprint's limit from a tiny shift in atoms' energy levels, 1 in 100, if its assumption holds.
SN 1987A antineutrinos and light: 2–5 × 10⁻³, resting on an absolute Shapiro delay that general relativity does not define on its own.Supernova 1987A's antineutrinos and light: 2 to 5 in 1,000, best read as an illustration.
Fischler, Lykken and Roberts's reading of the indirect evidence: 10⁻⁴, perhaps 10⁻⁸ (a remark, not a test).Three physicists' 2008 reading of the indirect clues: 1 in 10,000, perhaps 1 in 100 million (a remark, not a test).
GBAR's gravitational quantum states: a goal of about 10⁻⁵.GBAR's anti-atoms bouncing just above a surface: a goal of about 1 in 100,000.
Atom-interferometer fountains: a goal of about 10⁻⁶.Fountains of anti-atoms: a goal of about 1 in a million.
Alves et al., nuclear self-energy: 10⁻⁷, if their assumption holds; BASE with an assumed background potential: 1.8 × 10⁻⁷; where one theorist, Diego Blas, says it gets interesting: about 10⁻⁷.The 2009 preprint, from virtual pairs in atoms: 1 in 10 million, if its assumption holds; BASE with an assumed background gravity: 1.8 in 10 million; where one theorist says it gets interesting: about 1 in 10 million.
Alves et al., sea antiquarks: 10⁻⁹; this page's MICROSCOPE rescaling: 1–3 × 10⁻⁹ (both under Alves et al.'s assumption).The 2009 preprint, from the antiquarks inside protons: 1 in a billion; this page's rerun with the satellite: 1 to 3 in a billion (both under the preprint's assumption).
Eöt-Wash torsion balances, ordinary matter only: about 2 × 10⁻¹³.Twisting balances, ordinary matter only: about 2 in 10 million million.
MICROSCOPE, ordinary matter only: about 3 × 10⁻¹⁵.The MICROSCOPE satellite, ordinary matter only: about 3 in a million billion.

Read about a rung

A difference hiding at

Detect means at least 3σ (this page's rule).'Spot' means at least three times the test's error bar (this page's rule).

Direct drops precise enough to detect it (3σ)Real drops that could spot it

Planned drops that could detect it (3σ)Planned drops that could

Indirect arguments that already exclude it, if their assumptions hold (two are clock tests)Clues that already rule it out (each with an if)

As big as the change in gLike gravity's change

Precision of tests of antimatter's free fall, as |Δa|/g on a logarithmic scale. Rungs mix 1σ uncertainties (ALPHA-g 2023: ±0.21 g, this page's combination), bounds (ALPHA 2013: −65 < F < 110; indirect limits) and goals (the collaborations' stated aims, as of October 2026), so they compare only roughly. "Could detect" means a difference at least three times a test's 1σ precision or goal (this page's rule); a bound "excludes" anything larger than itself. BASE's two rungs are clock (redshift) tests, which bear on free fall only through the energy-conservation link between the two (Nordtvedt 1975). Indirect rungs are model-dependent: Alves, Jankowiak and Saraswat (2009, preprint) assume any anomaly shows up equally in virtual-pair and sea-antiquark energy; BASE's absolute limit assumes a background potential; SN 1987A's rests on an absolute Shapiro delay, which general relativity does not define on its own (Minazzoli et al. 2019); the MICROSCOPE rescaling is this page's arithmetic under Alves et al.'s assumption. Ordinary-matter tests (MICROSCOPE 2022; Eöt-Wash) are shown for scale only. Height anchor: h = R(1/√(1 − x) − 1), R = 6,371 km (this page's arithmetic).A ladder of how sharp each test is: each mark down the side is ten times sharper than the one above. Slide a hidden difference up and down: the readouts say which real drops could spot it, which planned ones might, and which indirect clues already rule it out, each with its own 'if'. The rungs mix error bars, limits and goals, so compare them only roughly.

The ladder makes the gap plain. ALPHA-g's error bar is as large as the weakening of g about 800 km above the ground; the 10⁻⁷ indirect limit is like the change from the floor to the top of a 32 cm step stool (this page's arithmetic).

Chapter 14Always a long shot

Five things 0.75 g does not mean:

Antimatter falls slower: 0.75 ± 0.21 g is consistent with 1 g; the gap is about the size of the error bar (Chapters 1 and 5).
Antimatter has negative mass: Its inertial mass and energy are positive and equal to matter's; only its gravitational mass was open (Chapter 2).
The equivalence principle is proved for antimatter: It was tested at about 20%. The indirect limits are far tighter, and each carries an assumption (Chapters 8 and 9).
CPT forced the answer: CPT relates antimatter to an anti-Earth; how antimatter falls on our Earth had to be measured (Chapter 11).
The missing antimatter is explained: The measurement does not touch the matter–antimatter asymmetry of the universe (Chapter 12).

Where things stand, as of 25 October 2026: ALPHA-g's 0.75 ± 0.13 ± 0.16 g is the only direct observation of gravity's effect on antimatter. ALPHA expects 1% or better; AEgIS and GBAR aim at 1% and have no result yet; the Antimatter Factory is off for the long shutdown. The arguments that made "up" a long shot have been joined by an observation; the arguments against a small difference remain arguments, each with its "if".

Antimatter falls down. The question now is whether it falls at exactly 1 g.

Glossary

Antimatter: Particles with the masses of ordinary ones and opposite charges, such as the positron and the antiproton.
Antihydrogen: An antiproton with a positron bound to it: the simplest anti-atom, and neutral.
Annihilation: A particle and its antiparticle meeting and turning into other particles or light, carrying away both masses' energy.
g: The acceleration of free fall at the Earth's surface; this page uses ALPHA's 9.81 m/s² (standard gravity is 9.80665 m/s²).
Inertial and gravitational mass: How hard a body resists being pushed, and how hard gravity pulls on it. For antimatter only the second was untested.
Weak equivalence principle: The rule that all bodies fall alike in the same gravitational field, whatever they are made of.
Eötvös parameter, η: The fractional difference between the free-fall accelerations of two materials; zero if the weak equivalence principle holds.
Magnetic trap: Fields that hold atoms by their magnetism, with mirror coils as lids and octupole windings round the sides.
Bias: A deliberate magnetic tilt between the trap's lids, quoted in units of g: a bias of 1 g pushes as hard as gravity.
Millikelvin as an energy: Small energies written as temperatures: 1 mK means k_B × 10⁻³ K, about 86 nano-electronvolts.
Uncertainty, σ, quadrature: A 1σ uncertainty is one standard deviation. Independent uncertainties combine as the square root of the sum of their squares.
Gravitational redshift: The change in light's frequency as it climbs or falls: 1.09 × 10⁻¹⁶ per metre near the Earth's surface.
Virtual pairs: Short-lived particle–antiparticle pairs that alter an atom's energy (vacuum polarisation); not real particles inside it.
C, P, T and the CPT theorem: Swapping matter and antimatter, mirroring space, reversing time. The weak force breaks each of them; all three together have never been seen broken.
Long Shutdown 3: CERN's accelerator shutdown that began in 2026; the Antimatter Factory switched off on 31 August 2026.

Sources

E. K. Anderson et al. (ALPHA Collaboration), "Observation of the effect of gravity on the motion of antimatter," Nature 621, 716 (2023), doi:10.1038/s41586-023-06527-1.
CERN, "ALPHA experiment at CERN observes the influence of gravity on antimatter" (press release, 27 September 2023), home.cern.
CERN Courier, "ALPHA-g clocks the freefall of antihydrogen" (3 November 2023), cerncourier.com.
P. Morrison, "Approximate nature of physical symmetries," Am. J. Phys. 26, 358 (1958), doi:10.1119/1.1996159.
L. I. Schiff, "Sign of the gravitational mass of a positron," Phys. Rev. Lett. 1, 254 (1958), doi:10.1103/PhysRevLett.1.254.
L. I. Schiff, "Gravitational properties of antimatter," Proc. Natl. Acad. Sci. USA 45, 69 (1959), doi:10.1073/pnas.45.1.69.
M. L. Good, "K₂⁰ and the equivalence principle," Phys. Rev. 121, 311 (1961), doi:10.1103/PhysRev.121.311.
M. M. Nieto and T. Goldman, "The arguments against 'antigravity' and the gravitational acceleration of antimatter," Phys. Rep. 205, 221 (1991), doi:10.1016/0370-1573(91)90138-c.
P. A. M. Dirac, "The quantum theory of the electron," Proc. R. Soc. A 117, 610 (1928), doi:10.1098/rspa.1928.0023.
P. A. M. Dirac, "Quantised singularities in the electromagnetic field," Proc. R. Soc. A 133, 60 (1931), doi:10.1098/rspa.1931.0130.
C. D. Anderson, "The apparent existence of easily deflectable positives," Science 76, 238 (1932), doi:10.1126/science.76.1967.238.
C. D. Anderson, "The positive electron," Phys. Rev. 43, 491 (1933), doi:10.1103/PhysRev.43.491.
O. Chamberlain, E. Segrè, C. Wiegand and T. Ypsilantis, "Observation of antiprotons," Phys. Rev. 100, 947 (1955), doi:10.1103/PhysRev.100.947.
G. Baur et al., "Production of antihydrogen," Phys. Lett. B 368, 251 (1996), doi:10.1016/0370-2693(96)00005-6.
M. Amoretti et al. (ATHENA), "Production and detection of cold antihydrogen atoms," Nature 419, 456 (2002), doi:10.1038/nature01096.
G. B. Andresen et al. (ALPHA), "Trapped antihydrogen," Nature 468, 673 (2010), doi:10.1038/nature09610.
C. J. Baker et al. (ALPHA), "Laser cooling of antihydrogen atoms," Nature 592, 35 (2021), doi:10.1038/s41586-021-03289-6.
M. J. Borchert et al. (BASE), "A 16-parts-per-trillion measurement of the antiproton-to-proton charge–mass ratio," Nature 601, 53 (2022), doi:10.1038/s41586-021-04203-w.
F. C. Witteborn and W. M. Fairbank, "Experimental comparison of the gravitational force on freely falling electrons and metallic electrons," Phys. Rev. Lett. 19, 1049 (1967), doi:10.1103/PhysRevLett.19.1049.
W. M. Fairbank, F. C. Witteborn and L. V. Knight, "An experiment to determine the gravitational force on free electrons and positrons," Science 144, 562 (1964), doi:10.1126/science.144.3618.562-b.
F. C. Witteborn and W. M. Fairbank, "Experiments to determine the force of gravity on single electrons and positrons," Nature 220, 436 (1968), doi:10.1038/220436a0.
ALPHA Collaboration and A. E. Charman, "Description and first application of a new technique to measure the gravitational mass of antihydrogen," Nat. Commun. 4, 1785 (2013), doi:10.1038/ncomms2787.
NIST, "CODATA 2022 recommended values of the fundamental physical constants" (μ_B, μ_B/k_B, m_p, m_e, u), physics.nist.gov/cuu.
R. Ahemeh, R. Akbari et al. (ALPHA), "Diagnosing antihydrogen energy distributions using spatio-temporal annihilation patterns," New J. Phys. 28, 083204 (2026), doi:10.1088/1367-2630/ae9c66.
R. V. Pound and G. A. Rebka Jr., "Apparent weight of photons," Phys. Rev. Lett. 4, 337 (1960), doi:10.1103/PhysRevLett.4.337.
T. Bothwell et al., "Resolving the gravitational redshift across a millimetre-scale atomic sample," Nature 602, 420 (2022), doi:10.1038/s41586-021-04349-7.
M. Fischler, J. Lykken and T. Roberts, "Direct observation limits on antimatter gravitation" (Fermilab note FERMILAB-FN-0822-CD-T, 2008, not refereed), arXiv:0808.3929.
K. Nordtvedt, "Quantitative relationship between clock gravitational 'red-shift' violations and nonuniversality of free-fall rates in nonmetric theories of gravity," Phys. Rev. D 11, 245 (1975), doi:10.1103/PhysRevD.11.245.
J. Scherk, "Antigravity: a crazy idea?," Phys. Lett. B 88, 265 (1979), doi:10.1016/0370-2693(79)90463-5.
G. Chardin and G. Manfredi, "Gravity, antimatter and the Dirac–Milne universe," Hyperfine Interact. 239, 45 (2018), doi:10.1007/s10751-018-1521-3, arXiv:1807.11198.
G. Chardin and J.-M. Rax, "CP violation: a matter of (anti)gravity?," Phys. Lett. B 282, 256 (1992), doi:10.1016/0370-2693(92)90510-B.
R. H. Price, "Negative mass can be positively amusing," Am. J. Phys. 61, 216 (1993), doi:10.1119/1.17293.
D. S. M. Alves, M. Jankowiak and P. Saraswat, "Experimental constraints on the free fall acceleration of antimatter" (2009, preprint), arXiv:0907.4110.
S. Schlamminger et al., "Test of the equivalence principle using a rotating torsion balance," Phys. Rev. Lett. 100, 041101 (2008), doi:10.1103/PhysRevLett.100.041101, arXiv:0712.0607.
T. A. Wagner, S. Schlamminger, J. H. Gundlach and E. G. Adelberger, "Torsion-balance tests of the weak equivalence principle," Class. Quantum Grav. 29, 184002 (2012), doi:10.1088/0264-9381/29/18/184002, arXiv:1207.2442.
P. Touboul et al. (MICROSCOPE), "MICROSCOPE mission: final results of the test of the equivalence principle," Phys. Rev. Lett. 129, 121102 (2022), doi:10.1103/PhysRevLett.129.121102, arXiv:2209.15487.
Particle Data Group (F. Takahashi et al.), "Review of Particle Physics," Int. J. Mod. Phys. A 41, 2630011 (2026), doi:10.1142/S0217751X26300115: K⁰ mass and K_L–K_S mass difference, pdg.lbl.gov.
T. Goldman, M. M. Nieto and V. D. Sandberg, "Kaons, quantum mechanics, and gravity," Mod. Phys. Lett. A 7, 3455 (1992), doi:10.1142/S021773239200286X.
R. J. Hughes and M. H. Holzscheiter, "Constraints on the gravitational properties of antiprotons and positrons from cyclotron-frequency measurements," Phys. Rev. Lett. 66, 854 (1991), doi:10.1103/PhysRevLett.66.854.
S. Ulmer et al. (BASE), "High-precision comparison of the antiproton-to-proton charge-to-mass ratio," Nature 524, 196 (2015), doi:10.1038/nature14861.
M. J. Longo, "New precision tests of the Einstein equivalence principle from SN1987A," Phys. Rev. Lett. 60, 173 (1988), doi:10.1103/PhysRevLett.60.173.
L. M. Krauss and S. Tremaine, "Test of the weak equivalence principle for neutrinos and photons," Phys. Rev. Lett. 60, 176 (1988), doi:10.1103/PhysRevLett.60.176.
O. Minazzoli, N. K. Johnson-McDaniel and M. Sakellariadou, "Shortcomings of Shapiro delay-based tests of the equivalence principle on cosmological scales," Phys. Rev. D 100, 104047 (2019), doi:10.1103/PhysRevD.100.104047, arXiv:1907.12453.
M. Villata, "CPT symmetry and antimatter gravity in general relativity," EPL 94, 20001 (2011), doi:10.1209/0295-5075/94/20001.
M. J. T. F. Cabbolet, "Comment to a paper of M. Villata on antigravity," Astrophys. Space Sci. 337, 5 (2012), doi:10.1007/s10509-011-0939-8.
A. Benoit-Lévy and G. Chardin, "Introducing the Dirac-Milne universe," Astron. Astrophys. 537, A78 (2012), doi:10.1051/0004-6361/201016103.
D. S. Hajdukovic, "Quantum vacuum and virtual gravitational dipoles: the solution to the dark energy problem?," Astrophys. Space Sci. 339, 1 (2012), doi:10.1007/s10509-012-0992-y.
M. Villata, "Antimatter gravity and the results of the ALPHA-g experiment," Ann. Phys. (Berlin) 536, 2300519 (2024), doi:10.1002/andp.202300519, arXiv:2503.03846.
D. S. Hajdukovic, "Antimatter gravity experiments, the astronomical challenges to ΛCDM cosmology and the quantum vacuum as a possible source of gravity in the universe," Int. J. Theor. Phys. 64, 93 (2025), doi:10.1007/s10773-025-05954-2, arXiv:2501.08368.
D. Hajdukovic, "The ALPHA-g experiment, the hunt for gravitational dipoles and the quantum vacuum as the source of gravity in the Universe" (2024, preprint; no journal reference as of 25 October 2026), arXiv:2403.15493.
G. M. Piacentino and A. Palladino, "Support for gravitationally-attractive composite antimatter and gravitationally-repulsive non-composite antimatter" (2025, condensed version for the Moriond Gravitation 2025 proceedings), arXiv:2504.03496.
C. S. Wu, E. Ambler, R. W. Hayward, D. D. Hoppes and R. P. Hudson, "Experimental test of parity conservation in beta decay," Phys. Rev. 105, 1413 (1957), doi:10.1103/PhysRev.105.1413.
J. H. Christenson, J. W. Cronin, V. L. Fitch and R. Turlay, "Evidence for the 2π decay of the K₂⁰ meson," Phys. Rev. Lett. 13, 138 (1964), doi:10.1103/PhysRevLett.13.138.
A. Angelopoulos et al. (CPLEAR), "First direct observation of time-reversal non-invariance in the neutral-kaon system," Phys. Lett. B 444, 43 (1998), doi:10.1016/S0370-2693(98)01356-2 (its reading as direct T violation was debated).
J. P. Lees et al. (BaBar), "Observation of time-reversal violation in the B⁰ meson system," Phys. Rev. Lett. 109, 211801 (2012), doi:10.1103/PhysRevLett.109.211801, arXiv:1207.5832.
M. Ahmadi et al. (ALPHA), "Characterization of the 1S–2S transition in antihydrogen," Nature 557, 71 (2018), doi:10.1038/s41586-018-0017-2.
R. Akbari et al. (ALPHA), "Four ppm measurement of the antihydrogen ground-state hyperfine splitting," Nature 653, 1022 (2026), doi:10.1038/s41586-026-10556-x.
C. Smorra et al. (BASE), "A parts-per-billion measurement of the antiproton magnetic moment," Nature 550, 371 (2017), doi:10.1038/nature24048.
M. Hori et al. (ASACUSA), "Buffer-gas cooling of antiprotonic helium to 1.5 to 1.7 K, and antiproton-to-electron mass ratio," Science 354, 610 (2016), doi:10.1126/science.aaf6702.
V. A. Kostelecký and J. D. Tasson, "Matter-gravity couplings and Lorentz violation," Phys. Rev. D 83, 016013 (2011), doi:10.1103/PhysRevD.83.016013, arXiv:1006.4106.
R. Caravita et al., "CERN AD/ELENA Antimatter Program" (2025, input to the European Strategy for Particle Physics update; not refereed), arXiv:2503.22471.
R. Akbari et al. (ALPHA), "Be⁺ assisted, simultaneous confinement of more than 15000 antihydrogen atoms," Nat. Commun. 16, 10106 (2025), doi:10.1038/s41467-025-65085-4.
CERN, "Breakthrough in antimatter production" (news, 18 November 2025), home.cern.
A. Powell (CERN), invited talk on ECR magnetometry of the ALPHA-g fields, CAP Congress 2026, Ottawa (24 June 2026; talk, work in progress), indico.global/event/15553.
D. Duque (TRIUMF), talk on rTPC position- and rate-dependent efficiencies, WNPPC 2026 (13 February 2026; talk, work in progress); F. Ghaly, "Towards demonstrating magnetic trapping of hydrogen in the HAICU experiment," WNPPC 2026 (13 February 2026; talk); and the WNPPC 2026 book of abstracts, indico.triumf.ca/event/830.
A. Powell, Magnetic Field Characterisation for Gravitational Free Fall Measurements of Antihydrogen in the ALPHA-g Experiment, Springer Theses (Springer, Cham, 2025), doi:10.1007/978-3-032-07190-3.
I. Carli (for ALPHA), "Status and performance of the ALPHA-g detectors," Nucl. Instrum. Meth. A 1080, 170676 (2025), doi:10.1016/j.nima.2025.170676.
INSPIRE-HEP, records citing Anderson et al. 2023 (record 2703716; 110 citing papers, accessed 25 October 2026), inspirehep.net.
CERN, "AEgIS" (experiment page), home.cern.
C. Amsler et al. (AEgIS), "Pulsed production of antihydrogen," Commun. Phys. 4, 19 (2021), doi:10.1038/s42005-020-00494-z.
L. T. Glöggler et al. (AEgIS), "Positronium laser cooling via the 1³S–2³P transition with a broadband laser pulse," Phys. Rev. Lett. 132, 083402 (2024), doi:10.1103/PhysRevLett.132.083402, arXiv:2310.08760.
R. C. Ferguson (for the AEḡIS Collaboration), "Probing the weak equivalence principle with a moiré deflectometer using antihydrogen at AEḡIS," J. Phys. Conf. Ser. 3149, 012004 (2025), doi:10.1088/1742-6596/3149/1/012004 (status paper).
CERN, "GBAR joins the anticlub" (news, 14 July 2023), home.cern.
P. Blumer (for GBAR), "Status and plans of the Gravitational Behaviour of Antihydrogen at Rest experiment" (Moriond 2022 contribution), arXiv:2203.16174.
P. Adrich et al. (GBAR), "Production of antihydrogen atoms by 6 keV antiprotons through a positronium cloud," Eur. Phys. J. C 83, 1004 (2023), doi:10.1140/epjc/s10052-023-12137-y.
GBAR Collaboration, "First measurement of the antihydrogen production cross section through the charge-exchange reaction of low-energy antiprotons with orthopositronium" (27 September 2026, preprint), arXiv:2609.33990.
J. Zhang et al., "Synthesis of a high intensity, superthermal muonium beam for gravity and laser spectroscopy experiments" (2025, preprint), arXiv:2512.19923.
B. Mikulec, "Accelerator Report: Final countdown – CERN's injector complex enters its last stretch before LS3" (CERN, 20 August 2026), home.cern.
A. Schaeffer, "All CERN accelerators now on the road to HiLumi" (CERN, 31 August 2026), home.cern.