betaFeedback

WorksPhysics and Astronomy

The speed limit of gravitational entanglement and the cost of noise

Abstract

Is gravity quantum? A proposed test lets two small masses, each in a superposition of two places, become entangled through their mutual gravity. Gravity is weak, so the masses must stay coherent for seconds while random forces disturb them. Here we ask how fast gravity can entangle two masses, and what noise costs. Without noise we find the maximum rate exactly: it falls as the inverse cube of the separation, and two positions per mass achieve it. Above a noise level set by the masses and their separation, gravity's leading pull can be imitated by classical signals, and every strategy is limited to a rate falling at least as the inverse fourth power. Exact, continuous error correction shields the masses from noise but hides them from each other: the best rate falls as the inverse seventh power, and six positions on a hexagon come within 10% of it at large separation.

Introduction

Nobody knows whether gravity obeys the rules of quantum mechanics. Every other force of nature does, but gravity is so feeble between small objects that its quantum side has never been seen. A way to look for it without a particle accelerator was proposed a few years ago[1, 2]. Take two small masses and put each of them in a superposition of two places: in two places at once, as quantum mechanics allows. If gravity can then entangle the two masses, so that their joint state cannot be described by a separate pure state for each mass, the gravitational pull between them cannot be something that two separate laboratories could imitate with ordinary measurements and classical messages. Many regard this as a decisive tabletop test, and it has become the goal of a growing experimental effort[3, 4, 5]. Related witnesses quantify the nonclassicality of mediated interactions through the correlations they generate[6].

The mechanism is easy to picture (Figure 1a). With each mass in two places at once, there are four combinations of positions, and the two masses are a different distance apart in each. Gravity is stronger when they are closer, so each combination picks up a slightly different quantum phase, a kind of internal clock reading. Wait long enough, and the four clocks drift apart in a way that leaves the two masses entangled.

Figure 1. The experiment, the obstacle and three speed limits. a, Two masses, each in two places at once inside its own region of radius rr; the regions are a distance dd apart. The four combinations of positions put the masses at four distances, gravity acts differently on each, and the masses become entangled. b, Random kicks from the surroundings and the pull of the partner mass both act on the position of a mass. A scheme that hides the position from the kicks hides it from gravity too. c, The results of this paper, as a sketch. Without noise, the fastest possible entanglement falls as 1/d31/d^{3}. Under strong random forces, no strategy can beat a rate that falls as 1/d41/d^{4}. With exact error correction, the best rate falls as 1/d71/d^{7}, and a six-position code nearly reaches it.

“Long enough” is the problem. Gravity has been measured between gold spheres two millimetres across[7], and molecules of tens of thousands of atomic mass units have been put in two places at once[8], but the entanglement test needs both at the same time: objects heavy enough to gravitate, in superpositions wide enough for gravity to tell the branches apart. For the masses usually discussed, specks a couple of micrometres across held a few hair-widths from each other, the gravitational effect takes many seconds to build up. All the while, stray gas molecules, electric fields and vibrations give the masses small random kicks, and a random kick is exactly what destroys a superposition of two places.

There is a well-known cure for noise in quantum devices, called error correction. One stores the quantum state not in two positions but in a cleverly chosen pattern of several positions, a code. The information then sits in the shape of the pattern, not in where the mass is, so that a random kick shifts the mass without learning or spoiling what is stored, and its damage can be undone. Codes of this kind are used in quantum sensing[9, 10, 11, 12] and were recently proposed for gravitational entanglement by Wang, Pedernales and Plenio[13]. But here the cure has a side effect (Figure 1b). A random kick and the pull of the partner mass grip the mass by the same handle: its position. A code that hides position from the noise hides it from gravity as well. Wang and colleagues showed that exact correction cancels the leading gravitational effect and found codes that keep a weaker, shorter-ranged remnant[13]. A related fact was known for unprotected masses: if the random forces are strong enough, the leading gravitational effect can be imitated by classical means and creates no entanglement[14, 15].

These observations raise three plain questions, and this paper answers them. How fast can gravity entangle two masses when nothing disturbs them? How much of that speed does noise take away, whatever one does about it? And if one shields the masses by error correction, how much speed can the best possible shield keep? The answers are three speed limits (Figure 1c). Each is stated for the complete Newtonian pull between two masses that are free to move inside two separate regions, and each holds for every strategy of the kind it describes, including ones nobody has thought of.

Results

A speed limit for gravity

We begin with the ideal case of no noise at all. Two masses, mAm_A and mBm_B, are each free to be anywhere inside a spherical region of radius rr; the centres of the two regions are a distance d>2rd > 2r apart (Figure 1a). The only thing that connects the two laboratories is Newton’s gravitational energy,

V(a,b)=−GmAmB∣dez+b−a∣(1)V(\mathbf{a},\mathbf{b})=-\frac{Gm_A m_B}{\lvert d\mathbf{e}_z+\mathbf{b}-\mathbf{a}\rvert} \tag*{(1)}

where a\mathbf{a} and b\mathbf{b} are the positions of the masses relative to the centres of their regions and ez\mathbf{e}_z points from one centre to the other. Everything else is allowed and costs nothing. Each laboratory may prepare any quantum state of its mass, use extra quantum systems as memories, measure, and send classical messages to the other. The two laboratories start out unentangled.

We measure entanglement in Bell pairs. A Bell pair is the basic unit of entanglement, as the bit is the basic unit of information: two quantum bits, one on each side, as entangled as two quantum bits can be. The entangling capacity is the largest number of Bell pairs per second that the laboratories can produce, averaged over a long experiment in which many pairs are made[16]. It is a speed limit: no protocol, however elaborate, entangles faster.

Two quantities set the scale of everything that follows. One is a rate, g0=GmAmB/(ℏd)g_0=Gm_A m_B/(\hbar d): the gravitational energy of the pair, turned into a frequency by dividing by Planck’s constant ℏ\hbar. For the masses we have in mind

it is a fraction of one per second. The other is the ratio u=r/du=r/d of the size of the regions to their separation, a number below one half. Our first result is the exact capacity. Whenever the regions are not too close, d≥(1+5)r≈3.24rd\geq(1+\sqrt{5})r\approx3.24r,

CG=βg02u21−4u2,β=1.9123…(2)C_G=\beta g_0\frac{2u^2}{1-4u^2},\qquad\beta=1.9123\ldots \tag*{(2)}

where β\beta is a constant known from the theory of the simplest interaction between two quantum bits[17, 18, 19].

The remarkable part is how little is needed. The best strategy is the simplest one: put each mass in two places at once, as two narrow wavepackets (well-localized lumps of the mass’s quantum wave) at the two ends of the diameter that points to the partner. Their four combinations sit at distances d−2rd-2r, d+2rd+2r, dd, and dd, and the mismatch between these distances is all the entangling power there is. Superpositions of many positions, motion sideways, entanglement with local memories: none of it beats two points on a line. That two points reach this rate was known; what is new is the proof that nothing else does. At large separation the speed limit falls as the cube of the distance, CG≈2βGmAmBr2/(ℏd3)C_G\approx2\beta Gm_A m_Br^2/(\hbar d^3).

To make this concrete, use one example throughout: two masses of 10−14 kg10^{-14}\,\mathrm{kg} in regions of radius r=125 μmr=125\,\mu\mathrm{m} whose centres are d=450 μmd=450\,\mu\mathrm{m} apart. Then Eq. (2) gives 0.060 Bell pairs per second. At the very best, gravity delivers one Bell pair every 17 seconds.

Why can cleverness not help? In the proof (Methods), every possible position of a mass is assigned a point in an abstract space, in such a way that the gravitational energy of two positions is given by how their points line up[20]. The entangling power available to any quantum state is limited by how far apart these points can be[21, 22]. All of them lie inside a ball, and the two ends of the axial diameter sit at opposite ends of it. Nothing can be farther apart than that, which is why two positions are enough. Two caveats: the capacity assumes ideal control with no limit on energy, and a fixed kinetic-energy budget lowers it in the collinear equal-mass model; and when the two regions nearly touch, a third position helps (Supplementary Sections S2 and S8).

Enough noise closes gravity's main channel

Now let the random kicks in. We model them in the simplest way: as a random force that is the same everywhere in a region and has no memory from one moment to the next. The forces on the two masses are independent, and nobody keeps a record of them. Such noise is described by one number, its strength Λ\Lambda, and what it does is easy to state: a mass in two places a distance LL apart loses its quantum coherence at the rate ΛL2\Lambda L^2. Wider superpositions are more fragile, by the square of their width. (A force with two-sided spectral density SFS_F per Cartesian component has Λ=SF/2ℏ2\Lambda= S_F/2\hbar^2.)

Gravity’s leading effect depends on size in the same way. When the masses are far apart compared with the width of their superpositions, the gravitational energy that distinguishes the four combinations of positions is proportional to the product of the two widths: make each superposition LL wide, and the entangling effect grows as L2L^2. Gravity and noise scale alike, the size of the superposition cancels out, and the contest between them is decided by a threshold that involves nothing but the two masses and their separation,

Λc=GmAmBℏd3.(3)\Lambda_c = \frac{Gm_A m_B}{\hbar d^3}. \tag*{(3)}

What happens at this threshold is striking (Figure 2a). Gravity’s leading effect, the one behind the 1/d31/d^3 law, couples the position of one mass to the position of the other. When the noise on the two masses satisfies ΛAΛB≥Λc\sqrt{\Lambda_A\Lambda_B} \ge\Lambda_c, this coupling together with the noise can be imitated exactly by a mundane procedure: each laboratory keeps measuring where its mass is, sends the result to the other as an ordinary, classical message, and the other gives its mass a matching push[14, 15]. Laboratories that only measure and exchange classical messages can never create entanglement; that is what makes entanglement a quantum resource. So above the threshold gravity’s main channel is closed: it is still there, but nothing quantum can pass through it.

Figure 2. Enough noise closes gravity’s main channel. a, Above a threshold of noise, gravity’s leading pull together with the random kicks is indistinguishable from two laboratories that measure the positions of their masses, exchange the results as classical messages and push their masses accordingly. Such a procedure cannot create entanglement. b, How fast entanglement can be made (vertical axis, as a fraction of the noiseless speed limit CGC_G) against the strength of the noise (horizontal axis, in units of the threshold Λc\Lambda_c of Eq. (3)), for regions ten radii apart. Grey: the limit for every strategy; the shaded area is impossible. The limit drops as the noise approaches the threshold, because the main channel closes and only the weaker effects of the potential are left (Eq. (4) and Methods). Blue: a rate achievable with two unprotected positions per mass; noise slows it, and beyond the marked point this scheme produces no entanglement at all. Orange: the six-position error-correcting code described below, which is immune to this noise but slow. Above the threshold, the best possible rate lies between the orange and grey lines. c, The noise at the threshold in laboratory units (newtons per root hertz; smaller is quieter), against separation, for three masses. Above each line the main channel is closed. The dot marks the running example.

This much was known for the leading effect alone. The real potential has further, weaker effects, and experimenters have further options: codes, feedback, memories, keeping a record of which errors occurred. Our second result covers all of it. Above the threshold, whatever the two laboratories do, the capacity of the complete Newtonian interaction obeys

Cnoise≤βg0B(u),C_{\mathrm{noise}} \le\beta g_0 B(u),
B(u)≈92(rd)3(u→0).(4)B(u) \approx\frac{9}{2}\left(\frac{r}{d}\right)^3 \qquad(u \to0). \tag*{(4)}

with the complete expression given in Methods. The rate now falls at least as fast as 1/d41/d^4: no ingenuity brings back the 1/d31/d^3 law. Compared with the noiseless speed limit, the most that any strategy can reach is 54% at d=6rd=6r and 28% at d=10rd=10r, and the percentage keeps falling in proportion to r/dr/d (Figure 2b). The idea of the proof is to split the true evolution into the part that measuring and messaging can imitate and a remainder, and to bound the entanglement that the remainder can produce while the laboratories do whatever they like in between (Methods and Supplementary Section S3).

Below the threshold the main channel is open, and two unprotected positions per mass do become entangled, although the noise slows them down (Figure 2b, blue). The threshold is not a knife edge. It marks where gravity’s leading effect stops working; the weaker effects let unprotected superpositions entangle slightly beyond it (Methods).

In laboratory terms, the threshold for the running example is Λc=6.9×105 m−2 s−1\Lambda_c = 6.9 \times10^5\ \mathrm{m}^{-2}\,\mathrm{s}^{-1}, which corresponds to a force noise of 1.2×10−5 N Hz−1/21.2 \times10^{-5}\ \mathrm{N\,Hz}^{-1/2} (Figure 2c): in a band of one hertz, fluctuations of less than a hundred-thousandth of the weight of a single hydrogen atom. The threshold rises with the product of the masses and falls with the cube of their separation, so heavier and closer masses can afford more noise.

A shield that hides too much

If the noise is too strong for unprotected masses, the natural response is to shield them with error correction. To see what a shield does to gravity, look at what a code must hide (Figure 3a).

Figure 3. A shield against noise also hides a mass from gravity. a, The two encoded states of a code (0 and 1) are patterns of positions. Without a shield they may differ in average position. To detect random kicks they must share their average position and can first differ in spread; to correct kicks they must share the spread as well and can first differ in their three-lobed shape. Gravity can use only what differs, so the interaction between the encoded states falls as 1/d31/d^3, 1/d51/d^5 and 1/d71/d^7. b, The six-position code: the encoded states are the two triangles of a hexagon, and the partner mass carries the mirror image. c, The fastest possible entanglement against separation; the labels give the slopes, that is, the powers of distance. Blue: the exact noiseless limit of Eq. (2) (dashed where its proof does not apply). Green and orange: codes that detect and codes that correct random kicks. Each band runs from an explicit code (four positions on a square, six on a hexagon) up to the limit that no code can exceed, so the best possible value lies inside the band. For correcting codes this is a rate that can be reached under noise; for detecting codes it is only the interaction between the encoded states. d, Why a hexagon beats a line. Gravity looks at a code from every direction ϕ\phi around the line joining the masses; the distance from the centre of the plot shows how strong a three-lobed signal it sees from each direction. The best code on a line shows 1/41/4 from all directions, the hexagon (1+3cos⁡2ϕ)/4(1 + 3\cos^2\phi)/4, and no code can show more than 0.713 on (root-mean-square) average (dashed).

A code stores one quantum bit in two states of a mass, the encoded states 0 and 1; each is a pattern of positions. Protection comes in two standard grades[23]. A code detects a random kick if one can always tell that a kick has happened without disturbing what is stored. For that, the two encoded states must have the same average position; otherwise the kick, which acts on position, would treat them differently. A code corrects a kick if the damage can also be undone. For that the two states must, in addition, have the same spread in every direction.

Gravity recognizes a mass by these same features. From a distance, a pattern of positions pulls on its partner first through its average position (physicists say its dipole), then through its spread (its quadrupole), then through its three-lobed shape (its octupole), and so on, each feature reaching less far than the one before. And a feature can help to entangle the two encoded bits only if it differs between the encoded states; a feature they share tells the partner nothing about which state is present.

Put the two together, and the consequence can be read off Figure 3a. Without a shield, gravity uses the average positions, and the encoded bits interact with a strength that falls as 1/d31/d^3. If kicks can be detected, the average positions are hidden; the first usable feature is the spread, and the interaction falls as 1/d51/d^5. If kicks can be corrected, averages and spreads are both hidden; only the three-lobed shapes remain, and the interaction falls as 1/d71/d^7. The shield and the price are one and the same thing, and each further grade of protection costs two more powers of r/dr/d. (Detection by itself does not remove the noise. It is the stepping stone that explains what correction costs.)

That the leading effect cancels under exact correction was shown by Wang and colleagues, who also found the faster fall-off in one family of codes[13]. The question we now answer is whether a better code could avoid the loss.

The best possible shield

It cannot. Take any code that corrects random kicks exactly. It may use any number of positions, store more than one quantum bit, and enlist extra quantum memories in its laboratory. Let the correction be applied as often as one likes, with the record of the errors discarded after each round, which is the simplest way to use a code. Our third result is that, at large separation, the best capacity CprotC_{\mathrm{prot}} of all such codes satisfies

29532≤Cprotβg0u6≤10.177583…+O(u).(5)\frac{295}{32} \le\frac{C_{\mathrm{prot}}}{\beta g_0 u^6} \le10.177583\ldots+ O(u). \tag*{(5)}

So the best shielded rate falls as 1/d71/d^7, four powers of distance faster than the unshielded limit, and the number in front is known to about 10%. A slightly weaker limit of the same form holds at every separation, not only far away (Figure 3c), and the pattern continues: codes that correct stronger classes of position errors lose four further powers of r/dr/d for each class (Methods).

The lower value in Eq. (5) comes from a code of appealing simplicity (Figure 3b). Take six positions at the corners of a regular hexagon, in a plane that contains the line between the masses, and let the two encoded states be the two triangles formed by alternate corners. Each triangle is the other one reflected through the centre. Such a reflection leaves averages and spreads unchanged and reverses every three-lobed feature. The two triangles therefore share their average position and their spread, as correction demands, and have opposite three-lobed shapes, which is all that gravity can still use. With the mirror-image hexagon at the partner mass, the two encoded bits interact with a strength of at least

ghex≥29532g0u6,(6)g_{\mathrm{hex}} \ge\frac{295}{32} g_0 u^6, \tag*{(6)}

and entangle at the rate βghex\beta g_{\mathrm{hex}}. This is not only a property of an idealized code: with smooth wavepackets and sufficiently frequent ideal correction, the rate is approached under the uniform random forces of the previous section, whatever their strength (Supplementary Section S4). Polygon arrangements of this kind have been studied before under the name of quantum spherical codes[24].

How good is the hexagon? Among all codes whose positions lie in a plane, however tilted and however much they store, 295/32295/32 is the best possible number in Eq. (5). Over all codes in three dimensions the number cannot exceed 10.18, so at large separation the hexagon delivers at least 90% of what any code could. Codes confined to the line between the masses, by contrast, cannot do better than 5/45/4 in place of 295/32295/32, a value reached by a code from quantum sensing[12]: the hexagon beats every one of them by a factor of 59/8≈7.459/8 \approx7.4.

This last number has a pleasing explanation (Figure 3d). Gravity inspects the three-lobed shape of a code from every direction around the line joining the masses, and the rate depends on the average of the square of what it finds. A code on the line looks the same from all sides and offers a signal of 1/41/4. The hexagon, seen from an angle ϕ\phi to its own plane, offers (1+3cos⁡2ϕ)/4(1 + 3\cos^2\phi)/4: four times as much face-on, and never less. Averaging the square gives exactly 59/859/8.

The statement “within 10%” concerns large separations. In the running example, where d=3.6 rd = 3.6\,r, the hexagon runs at 2.2% of the noiseless speed limit, while the general limit allows correcting codes up to 15.5%: there the hexagon is within a factor of seven of the best possible code, and the gap narrows to 10% only as the masses move apart. Exact correction with discarded error records is one strategy among many. Above the noise threshold it achieves a rate falling as 1/d71/d^7, while Eq. (4) forbids only rates above 1/d41/d^4. Whether some other strategy can do better in between is open.

The bill, in seconds

Entanglement has to be demonstrated, not just made, and the speed limits also say how long a demonstration must take. Suppose an experiment ends with a test. The test may raise a false alarm, reporting entanglement when there is none, with probability at most α\alpha; and it detects entanglement with probability pp when the experiment works as intended. We prove that the time for which the masses interact, averaged over many repetitions of the experiment, must satisfy

T‾≥1C[plog⁡2pα+(1−p)log⁡21−p1−α],(7)\overline{T} \ge\frac{1}{C}\left[p\log_2\frac{p}{\alpha} + (1-p)\log_2\frac{1-p}{1-\alpha}\right], \tag*{(7)}

where CC is the speed limit that applies (CGC_G without a shield). The bracket counts the bits of evidence that the test provides; for 1% false alarms and 95% detection it is 6.0 bits. The bound allows any measurement, feedback and stopping rule. Without a shield it is essentially the truth: as α\alpha becomes small, the least possible time approaches (p/CG)log⁡2(1/α)(p/C_G)\log_2(1/\alpha), and we give protocols that reach it.

Table I lists what this means for the running example. Without a shield, no experiment can make its case in less than 100 seconds of interaction, noise or no noise, and an explicit protocol with smooth finite-energy wavepackets and purely local measurements does it in 300 seconds (Supplementary Section S6). With a shield, every fixed exactly correcting code with discarded error records needs at least 648 seconds (Figure 4a). The hexagon’s coupling needs 1,119 seconds to turn two unentangled encoded bits into one Bell pair, where two unshielded positions need 25 seconds.

Figure 4. The bill. a, Least interaction time needed to demonstrate entanglement with 1% false alarms and 95% detection, for two masses of 10−1410^{-14} kg in regions of radius 125 μ\mum, against their separation. Lines are limits from Eq. (7): no experiment of the stated kind can be faster. Filled circles: the limits at 450 μ\mum. Open squares: explicit protocols, without a shield and without noise (300 s), and with a shield in the presence of noise (19.6 h). b, An imperfect shield. Horizontal axis: the calibrated correction imperfection δ\delta (Methods). Vertical axis: an upper bound on the coherent entangling rate retained by the code, in units of βg0\beta g_0. For a good shield the bound approaches the flat 1/d71/d^7 ceiling. Beyond δ≈(r/d)2\delta\approx(r/d)^2, its allowance for gravity’s main coupling (dotted lines) takes over.

Two rules of thumb summarize the bill. In time, shielding against random forces with the hexagon costs a factor that grows as (d/r)4(d/r)^4: 45 in the running example, about 2,200 at d=10rd = 10r. In mass, since every rate grows as the product of the masses, the lost speed is bought back by making both masses heavier by a factor that grows as (d/r)2(d/r)^2: 6.7 in the running example, 47 at d=10rd = 10r.

When the shield is imperfect

No real shield is perfect, infinitely fast, or faced only with uniform forces. Three further results show how the picture bends (Methods and Supplementary Section S7). First, a leaky shield leaks in both directions. If a code hides the average position and the spread only approximately, gravity’s main coupling leaks back in (Figure 4b). The 1/d71/d^7 law describes shields whose calibrated correction imperfection δ\delta (defined in Methods) is below about (r/d)2(r/d)^2. A leakier code keeps some of the stronger 1/d31/d^3 coupling, but in the codes we construct it also lets the noise back in.

Second, the kind of noise matters. A uniform random force is removed completely by the hexagon code. Noise that can tell the six positions apart, such as scattering events that reveal where the mass is, is not: it leaves a residual loss of coherence that no speed of correction removes, and for the continuously corrected hexagon, two equal logical superpositions become entangled only if the coupling exceeds half the residual rate on each mass, when those rates are equal. The residual loss belongs to the hexagon, not to the noise: we construct another family of codes that approaches a perfect Bell pair for any finite strength of the Gaussian localization noise considered here, at fixed geometry and bounded average kinetic energy, at the price of longer interaction times and ever finer control.

No shieldExact error correction
Limit: least interaction time, for any experiment100 s648 s (any code)
An explicit protocol300 s (twelve runs of 25 s; no noise)19.6 h (hexagon; 63 runs; noise that resolves the positions)
Time for the coupling to make one Bell pair25 s1,119 s (hexagon)

Table 1. The bill for the running example. Two masses of 10−1410^{-14} kg, r=125 μmr = 125\,\mu\mathrm{m}, d=450 μmd = 450\,\mu\mathrm{m}. The first two rows refer to a demonstration of entanglement with 1% false alarms and 95% detection (96% for the 19.6-hour protocol).

Third, correcting at a finite rate is a balancing act. Wait too long between corrections and errors pile up; correct too often and the imperfections of the corrections pile up instead. A fully worked example, with noise that resolves the positions, shows the scale of the task. About 28,000 corrections, one every 0.04 seconds and each with a complete-cycle error below one part in a million, yield a Bell pair of fidelity at least 0.817; a value above 0.5 certifies entanglement, and 1 is a perfect pair. Repeating this independently 63 times, 19.6 hours of interaction in

all, demonstrates entanglement with 1% false alarms and 96% detection, provided that the final readout is ideal and each run is prepared within trace distance 0.1% of its target. These figures are requirements on a future apparatus, not a claim that one exists.

Discussion

We set out with three questions and found three speed limits (Table II). Left undisturbed, gravity entangles two masses at a rate that is now known exactly; it falls as the cube of their separation, and two positions per mass are all it takes. Random forces above a threshold fixed by the masses and their separation close gravity’s main channel, and whatever is done about it the rate then falls at least as the fourth power. Shielding the masses by exact, continuously applied error correction removes the noise but hides the masses from each other, so that the rate falls as the seventh power, and six positions on a hexagon deliver, at large separation, nearly all that any such shield can.

One idea runs through all three. Gravity and random forces act on a mass through its position, so that whatever takes position away from the noise takes it away from gravity. Strong noise does this crudely, by scrambling the position until the leading gravitational effect could as well be a classical message. Error correction does it deliberately, by storing the quantum bit in features of the mass that a kick cannot reach, and gravity can barely reach them either. Seen this way, the limits are not peculiar to gravity. They use only the inverse-distance form of the interaction, and so apply equally to the electric force between two charges under the same model assumptions. Gravity is simply the case in which the interaction is so weak that the price decides whether an experiment is possible.

For those planning experiments, the results amount to a set of signposts. If the random forces are below the threshold of Eq. (3), no shield is required in principle, and the target is the noiseless limit of Eq. (2), reduced by the noise. If they are above it, the 1/d31/d^3 law is gone for good. Exact correction then offers a rate that is immune to the noise but smaller by roughly (r/d)4(r/d)^4, to be won back with heavier masses or with superpositions that are wider relative to their separation. The best design also changes shape: without a shield, the useful superposition lies along the line between the masses; with one, it spreads into a plane.

Our results rest on earlier ones. The imitation of the leading gravitational effect by measurement and messaging is due to Kafri, Taylor and Milburn[14], and its noise condition in general form to Li, Miki and Chen[15]; we extend it from the leading effect to the complete potential and to every strategy. The cancellation under exact correction and the recoverability of entanglement under noise are due to Wang, Pedernales and Plenio[13]; we add the limits that hold for all codes and the code that nearly reaches them. The codes themselves come from quantum sensing and from quantum spherical codes[12, 24], and larger superpositions and other geometries have been explored before[25, 26].

The limits also have edges, and it is as well to mark them. They concern Newton’s potential between masses confined to two regions, with ideal local control. The noise threshold assumes random forces that are uniform, memoryless and independent for the two masses. The shielded limit assumes fixed codes that correct exactly and continuously, with error records discarded. Within these assumptions the limits are proved; beyond them they are reference points. Noise with memory can change quantum-correlation dynamics and state-evolution speed limits[27].

The most interesting open question lies between two of our limits. Above the noise threshold, no strategy can beat 1/d41/d^4, and exact correction reaches 1/d71/d^7. Strategies that only detect errors and discard the runs in which one occurred, or that keep the error record, might do better;

SituationFastest possible entanglementWhat achieves it
No noiseexactly βg0 2u2/(1−4u2)\beta g_0\,2u^2/(1-4u^2); falls as 1/d31/d^3two positions per mass
Random forces above the threshold, any strategyat most βg0[92u3+…]\beta g_0[{\frac{9}{2}}u^3+\ldots]; falls at least as 1/d41/d^4open
Exact error correction, any codebetween 295/32295/32 and 10.1810.18 times βg0u6\beta g_0u^6; falls as 1/d71/d^7six positions on a hexagon (lower value)

caption: TABLE II. Three speed limits. Rates are in Bell pairs per unit time; g0=GmAmB/ℏdg_0=Gm_Am_B/\hbar d is the natural gravitational rate of the pair, u=r/du=r/d the ratio of region size to separation, and β=1.9123\beta=1.9123. The second and third rows show the behaviour at large separation; the complete expressions are in Methods.

deciding this means counting the failed attempts and following what happens to the runs that survive. Also open are the exact number in front of the shielded limit, which lies between 295/32295/32 and 10.1810.18, the capacity when the regions nearly touch, and the effect of a common limit on energy or on the speed of control. Above all, a physical way to carry out the correction on the motion of a mass, with the accuracy that our examples demand, has yet to be invented. Gravity’s pull between two specks of matter is the faintest thing anyone has proposed to entangle with. The limits found here say how quickly that can happen at best, and what it costs to listen for it through the noise.

Methods

References

  1. [1]S. Bose, A. Mazumdar, G. W. Morley, H. Ulbricht, M. Toroš, M. Paternostro, A. A. Geraci, P. F. Barker, M. S. Kim, and G. Milburn, Spin entanglement witness for quantum gravity, Physical Review Letters 119, 240401 (2017).DOI
  2. [2]C. Marletto and V. Vedral, Gravitationally induced entanglement between two massive particles is sufficient evidence of quantum effects in gravity, Physical Review Letters 119, 240402 (2017).DOI
  3. [3]D. Carney, P. C. E. Stamp, and J. M. Taylor, Tabletop experiments for quantum gravity: a user’s manual, Classical and Quantum Gravity 36, 034001 (2019).DOI
  4. [4]T. Krisnanda, G. Y. Tham, M. Paternostro, and T. Paterek, Observable quantum entanglement due to gravity, npj Quantum Information 6, 12 (2020).DOI
  5. [5]L. Lami, J. S. Pedernales, and M. B. Plenio, Testing the quantumness of gravity without entanglement, Physical Review X 14, 021022 (2024).
  6. [6]R. Ganardi, E. Panwar, M. Pandit, B. Wolonczewicz, and T. Paterek, Quantitative nonclassicality of mediated interactions, PRX Quantum 5, 010318 (2024).DOI
  7. [7]T. Westphal, H. Hepach, J. Pfaff, and M. Aspelmeyer, Measurement of gravitational coupling between millimetre-sized masses, Nature 591, 225 (2021).DOI
  8. [8]Y. Y. Fein, P. Geyer, P. Zwick, F. Kiałka, S. Pedalino, M. Mayor, S. Gerlich, and M. Arndt, Quantum superposition of molecules beyond 25 kDa, Nature Physics 15, 1242 (2019).DOI
  9. [9]S. Zhou, M. Zhang, J. Preskill, and L. Jiang, Achieving the Heisenberg limit in quantum metrology using quantum error correction, Nature Communications 9, 78 (2018), arXiv:1706.02445.DOI
  10. [10]R. Demkowicz-Dobrzanski, J. Czajkowski, and P. Sekatski, Adaptive quantum metrology under general Markovian noise, Physical Review X 7, 041009 (2017).
  11. [11]D. Layden and P. Cappellaro, Spatial noise filtering through error correction for quantum sensing, npj Quantum Information 4, 30 (2018).
  12. [12]D. Layden, S. Zhou, P. Cappellaro, and L. Jiang, Ancilla-free quantum error correction codes for quantum metrology, Physical Review Letters 122, 040502 (2019), arXiv:1811.01450.DOI
  13. [13]Y.-X. Wang, J. S. Pedernales, and M. B. Plenio, Recoverability of gravitational entanglement under decoherence (2026), arXiv:2605.04544v1.
  14. [14]D. Kafri, J. M. Taylor, and G. J. Milburn, A classical channel model for gravitational decoherence, New Journal of Physics 16, 065020 (2014).DOI
  15. [15]A. Li, D. Miki, and Y. Chen, Universal bound for entanglement generation (2026), arXiv:2605.26215.arxiv.org/abs/2605.26215
  16. [16]C. H. Bennett, A. W. Harrow, D. W. Leung, and J. A. Smolin, On the capacities of bipartite hamiltonians and unitary gates, IEEE Transactions on Information Theory 49, 1895 (2003), arXiv:quant-ph/0205057v4.DOI
  17. [17]W. Dür, G. Vidal, J. I. Cirac, N. Linden, and S. Popescu, Entanglement capabilities of nonlocal hamiltonians, Physical Review Letters 87, 137901 (2001), arXiv:quant-ph/0006034.
  18. [18]A. M. Childs, D. W. Leung, F. Verstraete, and G. Vidal, Asymptotic entanglement capacity of the Ising and anisotropic Heisenberg interactions, Quantum Information and Computation 3, 97 (2003), arXiv:quant-ph/0207052v2.DOI
  19. [19]A. M. Childs, D. W. Leung, and G. Vidal, Reversible simulation of bipartite product hamiltonians, IEEE Transactions on Information Theory 50, 1189 (2004), arXiv:quant-ph/0303097v1.DOI
  20. [20]R. L. Frank and E. H. Lieb, Inversion positivity and the sharp Hardy–Littlewood–Sobolev inequality, Calculus of Variations and Partial Differential Equations 39, 85 (2010), arXiv:0904.4275 [math.FA].arxiv.org/abs/0904.4275
  21. [21]S. Bravyi, Upper bounds on entangling rates of bipartite Hamiltonians, Physical Review A 76, 052319 (2007).DOI
  22. [22]K. Van Acoleyen, M. Mariën, and F. Verstraete, Entanglement rates and area laws, Physical Review Letters 111, 170501 (2013).DOI
  23. [23]E. Knill and R. Laflamme, Theory of quantum error-correcting codes, Physical Review A 55, 900 (1997).DOI
  24. [24]S. P. Jain, J. T. Iosue, A. Barg, and V. V. Albert, Quantum spherical codes, Nature Physics 20, 1300 (2024), arXiv:2302.11593.DOI
  25. [25]J. Tilly, R. J. Marshman, A. Mazumdar, and S. Bose, Qudits for witnessing quantum-gravity-induced entanglement of masses under decoherence, Physical Review A 104, 052416 (2021), arXiv:2101.08086 [quant-ph].DOI
  26. [26]H. C. Nguyen and F. Bernards, Entanglement dynamics of two mesoscopic objects with gravitational interaction, The European Physical Journal D 74, 69 (2020), arXiv:1906.11184 [quant-ph].arxiv.org/abs/1906.11184
  27. [27]K. G. Paulson, E. Panwar, S. Banerjee, and R. Srikanth, Hierarchy of quantum correlations under non-markovian dynamics, Quantum Information Processing 20, 141 (2021), arXiv:2004.11208 [quant-ph].DOI
  28. [28]H. M. Wiseman, Quantum theory of continuous feedback, Physical Review A 49, 2133 (1994).pubmed.ncbi.nlm.nih.gov/9910465
  29. [29]M. Horodecki, P. Horodecki, and R. Horodecki, Inseparable two spin-1/2 density matrices can be distilled to a singlet form, Physical Review Letters 78, 574 (1997), arXiv:quant-ph/9607009.
  30. [30]I. Devetak and A. Winter, Distillation of secret key and entanglement from quantum states, Proceedings of the Royal Society A 461, 207 (2005).
  31. [31]H. Zhu and M. Hayashi, Optimal verification and fidelity estimation of maximally entangled states, Physical Review A 99, 052346 (2019), arXiv:1901.09772.DOI
  32. [32]S. Pallister, N. Linden, and A. Montanaro, Optimal verification of entangled states with local measurements, Physical Review Letters 120, 170502 (2018), arXiv:1709.03353v3.DOI
  33. [33]A. S. Holevo, M. E. Shirokov, and R. F. Werner, Separability and entanglement-breaking in infinite dimensions (2005), arXiv:quant-ph/0504204.arxiv.org/abs/quant-ph/0504204
  34. [34]M. Ozawa, Quantum measuring processes of continuous observables, Journal of Mathematical Physics 25, 79 (1984).DOI
  35. [35]S. Friedland and L.-H. Lim, Nuclear norm of higher-order tensors, Mathematics of Computation 87, 1255 (2018).
  36. [36]C. H. Bennett, D. P. DiVincenzo, J. A. Smolin, and W. K. Wootters, Mixed-state entanglement and quantum error correction, Physical Review A 54, 3824 (1996), arXiv:quant-ph/9604024.DOI
  37. [37]C. Bény and O. Oreshkov, General conditions for approximate quantum error correction and near-optimal recovery channels, Physical Review Letters 104, 10501 (2010), arXiv:0907.5391 [quant-ph].DOI
  38. [38]S. Das, S. Bäuml, and M. M. Wilde, Entanglement and secret-key-agreement capacities of bipartite quantum interactions and read-only memory devices, Physical Review A 101, 012344 (2020), arXiv:1712.00827v3.DOI
  39. [39]A. W. Harrow and A. Lowe, Optimal quantum circuit cuts with application to clustered hamiltonian simulation, PRX Quantum 6, 010316 (2025), arXiv:2403.01018v2.DOI

Paper details

vlnc:W_p7hfmgv3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents