How entanglement is actually made

Laboratories make entangled light with a laser and a sliver of transparent crystal, from a fraction of a millimetre to a centimetre or two long [Reference 1,Reference 2]. Almost all of the laser light goes straight through. But every so often one particle of light, a photon, vanishes and two new ones appear in its place [Reference 3]. Catch those twins in the right directions and their polarizations (the direction in which each light wave wiggles, up-down or side-to-side) are entangled. Quantum theory gives neither photon a definite polarization until it is checked. Yet in the classic setup, check both for “up-down or side-to-side” and they come out opposite (ideally every time); check both along diagonal directions instead and their answers are still tightly linked, which is what lets such pairs beat the limits that ready-made answers would obey [Reference 4,Reference 2].

The process is called spontaneous parametric down-conversion (SPDC): “down-conversion” because one high-energy photon is converted into two lower-energy ones (“parametric” is a technical label you can safely ignore) [Reference 3]. This article follows one photon through the crystal and ends 500 km up, with the Chinese satellite Micius, which shared such pairs between two observatories 1,203 km apart [Reference 2]. What entanglement is (and isn’t) is the subject of our companion explainer. Testing it with a Bell experiment (a test that rules out pre-set answers) is a story for a future article.

A crystal that almost never splits light

In ordinary glass, photons slow down a little, bend at surfaces, and come out the same color. Some crystals behave more strangely. Their atoms are arranged with no center of symmetry: swap every point with the point directly opposite it through the center, and the crystal looks different. As a result, the electrons respond to a strong light wave in a lopsided way [Reference 6]. That lopsided response lets such a nonlinear crystal mix light: merge two photons into one of higher energy or, in reverse, split one photon into two [Reference 3].

The split happens on its own, with no second beam to trigger it, hence spontaneous. Pre-quantum (“classical”) physics can’t explain it: with no second beam, nothing in that picture would start the split. Quantum theory predicts it, and gives each incoming “pump” photon only a tiny chance of splitting [Reference 3].

How tiny? A review of the field puts the fraction of pump photons that split somewhere between one in a hundred thousand and one in a trillion, depending on the crystal and the setup [Reference 3]. Experiments collect only some of the pairs a crystal makes: those flying in the right directions with the right colors. The Micius satellite’s source, for example, delivered 5.9 million pairs per second into its collection optics from 30 milliwatts (mW, thousandths of a watt) of violet light with a wavelength of 405 nanometres (nm, billionths of a metre). That beam carries some 6 × 10¹⁶ photons per second, so about one photon in ten billion ended up as a collected pair; the crystal itself made more pairs than that [Reference 2].

Each split is a matter of chance, so pairs appear at unpredictable moments [Reference 3]. Experiments pick out the twins by coincidence counting: two detectors, and a record of only those clicks that arrive within a few billionths of a second of each other [Reference 3,Reference 2]. The faint glow from such splits, called parametric fluorescence, was already being reported in 1967 [Reference 7,Reference 8]. In 1970 David Burnham and Donald Weinberg showed that it comes in simultaneous pairs: shining an ultraviolet laser into a crystal of ammonium dihydrogen phosphate, they saw coincidences only when the two detectors sat where conservation of energy and momentum said the partners should go [Reference 9].

The energy budget

A photon’s energy is set by its wavelength: the shorter the wavelength, the more energy it carries. Physicists measure such small energies in electronvolts (eV), the energy an electron picks up when pushed by one volt. A 405 nm violet photon carries 3.06 eV and an 810 nm photon carries 1.53 eV, so a violet photon can split into two 810 nm photons with nothing left over.

But “two 810 nm photons” is only the average. The crystal doesn’t insist on an even split. A pair might be one 800 nm photon and one 820.25 nm photon: unequal shares, same total. (Not 820 nm, because energy goes with one divided by the wavelength, so wavelengths don’t split evenly.) Uneven pairs leave at slightly different angles, thickening the cones described below. Experiments put color filters in front of the detectors that pass only a narrow band of wavelengths: a band 5 nm wide in the classic 1995 experiment, 20 nm wide at the satellite’s ground stations [Reference 4,Reference 2,Reference 3].

The hard part: momentum

Momentum is the trickier bookkeeping. In empty space a photon’s momentum is also set by its wavelength, and a 405 nm photon carries exactly twice the momentum of an 810 nm one, so splitting it straight ahead into two 810 nm photons would balance perfectly. Inside a crystal, though, light slows down by a factor called the refractive index: an index of 1.69 means light’s wave crests advance 1.69 times slower than in vacuum. Slowed light has its crests packed closer together. That runs against school intuition, where a slower ball carries less momentum, but for light inside a material the momentum that must balance goes with how tightly the crests are packed, so it grows by the same factor [Reference 3,Reference 5].

Here’s the catch. In almost every transparent material, violet light is slowed more than infrared light [Reference 3]. In BBO, light with what is called the “ordinary” polarization (explained below) has an index of 1.692 at 405 nm but only 1.660 at 810 nm [Reference 10]. If those were the only indices on offer, the pump photon’s momentum would come out 1.9% larger than its two daughters’ combined.

That sounds small, but it is fatal. Pairs are made all along the crystal, and their light only builds up if it stays in step, like a crowd clapping in time. With a 1.9% mismatch, light made a little deeper in the crystal drifts out of step with light made earlier. After about 6 micrometres (μm, thousandths of a millimetre) it is fully out of step and starts cancelling what came before, so almost nothing comes out [Reference 5]. That distance is far shorter than any crystal in this story.

The way out, in BBO, is birefringence. In a birefringent crystal, how fast light travels depends on its polarization. For one polarization, called ordinary, the refractive index is the same in every direction. For the other, called extraordinary, the index changes with the angle between the light and a special direction in the crystal called the optic axis. In BBO the extraordinary index is the smaller one [Reference 3]. So by sending an extraordinary pump in at just the right angle, you can lower its index until the books balance.

This is phase matching: the condition that the new photons’ momenta add up exactly to the pump’s, so that light made throughout the crystal stays in step [Reference 3]. It comes in two flavors. In type-I, the two new photons have the same polarization. In type-II, they have perpendicular polarizations, one ordinary and one extraordinary [Reference 3,Reference 4]. For type-II, with a 405 nm pump and two 810 nm photons travelling straight ahead, our calculation with modern measurements of BBO gives a pump angle of 41.6° to the optic axis [Reference 10]. Crystals are cut, their faces sliced at a chosen angle to the internal structure, so that this angle is reached when the laser hits the face squarely, then fine-tuned by tilting [Reference 4].

Birefringence isn’t the only trick. In a periodically poled crystal, the crystal’s lopsidedness is manufactured to flip direction every few micrometres, right where the light would start cancelling itself. Each flip pulls the light back into step, so the split keeps building up without any special angle, and the light can even travel straight along one of the crystal’s natural axes [Reference 5]. One such material, periodically poled potassium titanyl phosphate (PPKTP), with its pattern repeating every 10.0 μm (one flip every 5 μm), has been used to split 405 nm photons into 810 nm pairs that leave together in a straight line, tuned by temperature rather than angle [Reference 11]. Micius carried this kind of crystal [Reference 2].

Two cones of light

Phase matching doesn’t pick a single direction. Send one daughter photon slightly up and its partner slightly down, and the momentum books can still balance: the sideways pushes cancel because the partners leave on opposite sides of the beam [Reference 4]. The allowed directions form a ring, so the photons fly out along cones. In type-II each cone sits off to one side of the beam rather than centered on it, and there are two of them: ordinary and extraordinary light are slowed by different amounts, so their allowed directions differ [Reference 4,Reference 3]. In the classic 1995 experiment the two cones were vertically and horizontally polarized, and every pair had one photon on each [Reference 4].

Tilting the crystal by a degree or so changes everything. At the straight-ahead cut (the angle at which pairs could leave exactly along the beam), each cone’s edge just reaches the beam, so the two cones touch there. Tilt one way and they pull apart [Reference 4]; about 1.4° from the touching point, our calculation shows, the equal-color cones shrink to two narrow beams and then vanish [Reference 12] (the crystal still makes unequal-color pairs, such as 782 nm plus 840 nm). Tilt the other way and the cones lean in toward the beam and cross along two lines [Reference 4], which show up as two points on a screen behind the crystal. In our calculation, a tilt of about 1.2° puts the crossings 6° apart. Try it below.

Where the cones cross

Look along one of the two crossing lines. A photon arriving there might be a horizontally polarized photon from one cone, or a vertically polarized photon from the other. Its partner, at the opposite crossing, is the reverse. If nothing in the experiment can tell these two possibilities apart, the pair is described by a superposition, a quantum combination of both at once: “horizontal here, vertical there” and “vertical here, horizontal there” [Reference 4]. That is polarization entanglement. Every pair is already entangled in energy and direction, because both photons share one pump photon’s energy and momentum [Reference 1]; the crossings add entanglement in polarization, which is far easier to work with in practice [Reference 4].

It is more than each photon secretly leaving with an opposite polarization. Pairs like that would show no link at all when checked along diagonal directions. Entangled pairs stay linked there too, and checking diagonal directions is how experimenters show that both possibilities are truly present at once [Reference 4]. Turning that link into a test that no pre-set answers can pass is what a Bell experiment does.

The crossing-cone source was demonstrated in a 1995 experiment by Paul Kwiat, Anton Zeilinger and colleagues in Innsbruck and Maryland, who note that such a source had also been suggested by A. Garuccio [Reference 4]. Nearly all earlier down-conversion experiments discarded half their detected pairs. The type-II crossings gave the entangled state directly, with no detected pairs discarded [Reference 4], though light away from the crossings goes unused [Reference 1]. Zeilinger later shared the 2022 Nobel Prize in Physics, whose citation begins “for experiments with entangled photons” [Reference 14].

Erasing the clues

There is a catch, and it comes from the same birefringence that made phase matching possible. Inside the crystal, horizontally and vertically polarized photons travel at different speeds and drift slightly sideways relative to each other [Reference 4,Reference 3]. Because one polarization is always the faster one, the order in which the two detectors fire could, in principle, reveal which cone each photon came from, and then the entanglement is lost [Reference 4].

Kwiat’s team put the delay in their 3 mm crystal at 385 femtoseconds (millionths of a billionth of a second). Each photon is a burst of wave about that long, so a delay that size could reveal which arrived first [Reference 4]. Their fix was a half-wave plate, a thin optic that swaps horizontal and vertical polarization, in each beam, followed by a second BBO crystal half as long as the first. Why half? A pair born at the entrance face crosses all 3 mm, so the faster polarization builds up its full head start; a pair born at the exit gets none. With the polarizations swapped, the 1.5 mm crystal takes half of that full head start back: the entrance pair ends half a head start ahead, the exit pair half behind. Every pair that would arrive “horizontal first” now has a twin, born the same distance on the other side of the crystal’s middle, that arrives “vertical first” by the same amount, so timing no longer gives the answer away. The same trick largely corrects the sideways drift too [Reference 4].

From 1995 to the classroom

The 1995 source was not the violet-laser setup shown above. The team pumped a 3 mm BBO crystal with ultraviolet light at 351.1 nm from an argon-ion laser, so their twin photons came out at 702 nm rather than 810 nm. The crystal was cut at 49.2° and tilted by 0.72°, which put the two crossing lines 6.0° apart [Reference 4].

It counted more than 1,500 entangled pairs per second, with a cleanliness score called visibility of 97.8 ± 1.0% (100% is perfect) [Reference 4]. The team violated a Bell inequality, a limit that pairs carrying pre-set answers can never beat, by more than 100 standard deviations, a measure of how far a result sits beyond what chance could easily produce (more on that here). Like all tests of its time, it relied on extra assumptions about the detectors and the source [Reference 4].

Four years later Kwiat and colleagues introduced a brighter design: two thin type-I crystals stacked face to face, one rotated a quarter turn relative to the other [Reference 1]. Each crystal splits only one pump polarization, and both photons of its pairs share one polarization, at right angles to the pump’s. Pump with light polarized at 45°, half-horizontal and half-vertical, and a pair could have come from either crystal. If the two cones overlap well enough to hide which, pairs all the way around the cone are expected to be entangled, as “both horizontal” plus “both vertical”: instead of two crossing points, the whole ring becomes usable, which is why the design is brighter [Reference 1]. Here the photons always match instead of disagreeing; both kinds of pair are entangled. The design was more than ten times brighter per unit of pump power, with a raw visibility of 98.8 ± 0.2% (99.6% after subtracting accidental coincidences, clicks from unrelated photons that happened to arrive together) [Reference 1].

With an affordable 405 nm violet diode laser, the two-crystal design became an undergraduate experiment. One 2002 classroom version used a 5 mW diode and two crystals 0.1 mm thick, and still violated a Bell inequality [Reference 6].

Splitting pairs from orbit

In 2017 a team led by Jian-Wei Pan reported sending entangled photons from Micius, a 635 kg satellite orbiting about 500 km up, to two ground stations: Delingha, in Qinghai province, and the Gaomeigu Observatory near Lijiang, in Yunnan, 1,203 km apart [Reference 2].

Its source used the same splitting physics but balanced momentum differently, with no BBO cones at all. A 405 nm laser diode pumped a 15 mm PPKTP crystal inside a loop of mirrors called a Sagnac interferometer [Reference 2]. The pump light was split by polarization and sent round the loop both ways at once, so a pair could be born from light going clockwise or anticlockwise. A polarizing beam splitter (an optic that sends horizontally and vertically polarized light in different directions) merged the two routes so that nothing reveals which, the same “hide the origin” trick as the crossing cones, giving polarization-entangled pairs at about 810 nm [Reference 2,Reference 11].

Why go to space? In fiber, every kilometre swallows the same fraction of the light, so losses multiply with distance. From orbit, most of the path is nearly empty space, where almost nothing absorbs the light; photons are still lost as the beam spreads and is pointed imperfectly, and to absorption and turbulence concentrated in the lowest 10 km or so of the atmosphere [Reference 2,Reference 15]. For standard fiber the authors estimate 10⁻¹⁷ pairs per second from a source making 10 million pairs per second: about one pair every three billion years, by our arithmetic. Even the best commercial fiber of the time would have been about a trillion times (12 orders of magnitude) less efficient than the satellite, and, the authors estimate, fiber at silica’s physical loss limit would still lose by 5 to 7 orders of magnitude [Reference 2]. Despite heavy losses on the way down, the stations recorded about one photon pair per second, roughly eight in nine of them real pairs rather than chance clicks [Reference 2].

The pairs arrived still entangled: the team estimated a fidelity (a score of how close the pairs were to perfect entanglement, where 1 is perfect) of at least about 0.87, with an uncertainty of 0.09 [Reference 2]. A Bell test gave a score of S = 2.37 ± 0.09, above the limit of 2 that any explanation based on pre-set instructions carried by each photon must obey, by about four standard deviations [Reference 2]. That is short of five, the traditional threshold for claiming a discovery in particle physics, as a preprint (a paper shared before formal peer review) by Louis Lyons discusses [Reference 16] (see our five-sigma explainer). But this was a check, at a new distance, of an effect laboratory Bell tests had already confirmed many times, some by more than 100 standard deviations [Reference 6,Reference 4]. In 2020 the same satellite delivered pairs to Delingha and Nanshan, 1,120 km apart, and the team used them for quantum key distribution, building a shared secret key at 0.12 bits per second [Reference 17].

Open questions

A 2018 review noted that the process’s low efficiency limits applications, and listed efforts to improve it using optical fibers, light-guiding channels etched in silicon, semiconductor chips and engineered “metamaterials” [Reference 3].

As of 2022, the Nobel committee reported that a satellite in a higher orbit was under construction that would distribute entangled pairs directly over 10,000 km [Reference 15].

Finally, every long-distance Bell test has to contend with photon loss and with assumptions about how measurement settings are chosen, the so-called loopholes. Which of those assumptions the Micius test needed is the subject of a future article [Reference 2].

Five Sigma, “How entanglement is actually made,” Five Sigma (12 October 2026), https://thefivesigma.com/articles/how-entanglement-is-made. Version of 12 October 2026.