Non-local gates are better in some architectures! If we assumed they are still on the same chip...

It goes something like this: "architecture X is better suited for non-local gates". Every time I read it, my DQC brain lights up: oh, nice, someone has a cheaper way of doing gates across QPUs? And every time, I read on, and no... They mean a two-qubit gate between two qubits that aren't neighbours. On the same chip. In the same fridge, or the same vacuum chamber.

That's not a distributed gate... That's a gate! Why!?!?

Two fields, one word

In DQC, a non-local gate has a fairly precise meaning (see the hypergraph post): an entangling gate acting on qubits that live in different QPUs, implemented by consuming entanglement distributed over a quantum network, through a communication primitive such as teleportation or a cat-entangler. It costs EPR pairs, it costs classical communication, and it costs whatever noise the link injects.

In the hardware and compiler world, "non-local" (or "long-range") is a more relaxed term, usually refering to a gate between two qubits that aren't directly coupled in the device's native connectivity graph. On a heavy-hex superconducting chip, that's anything beyond nearest neighbours. The cost is SWAPs, or shuttling time, or a special long-range coupler. Not a network.

This leads to a bit of a problem. A lot of architectures are known to have "better" non-local gates:

Neutral atoms: atoms can be physically moved around with optical tweezers, so "far away" qubits can be brought together before the gate (see e.g. Bluvstein et al.).

Trapped ions: within one chain, the shared motional modes give you something close to all-to-all connectivity. Across zones, QCCD architectures shuttle ions around (see Pino et al.).

Superconducting chips with long-range couplers: qLDPC codes like the bivariate bicycle codes (Bravyi et al.) need connections beyond a 2D grid, and the hardware roadmaps are adding couplers to provide them.

And this gets conflated with them being better for DQC.

And it's not like I want to argue that there is no correlation whatsoever between both definitions. And sure, words have multiple meanings, but this conflation is causing actual problems, at least in my corner of the world:

1. The literature search is polluted: Search for "non-local gates" and you get a mix of routing papers, neutral-atom papers and DQC papers. When you're writing a related-work section, telling them apart can take quite a while and be very confusing for new joiners.
2. Metrics don't transfer: A compiler paper that reports "number of non-local gates" might be counting inserted SWAPs, or it might be counting ebits consumed. These are very different units! One costs a few extra two-qubit gates and some depth. The other costs entanglement generation, which in today's networks is orders of magnitude slower and noisier than a local gate.
3. It sells the wrong story. "Architecture X is better at non-local gates" can be read by a non-specialist (a funder, say) as "architecture X makes distribution easier". It may or may not! Moving atoms inside a vacuum chamber tells you very little about how well those atoms interface with photons going down a fibre. And unfortunately the line has become so blurry that it's even starting to confuse experts.

My suggestion...

If you care about such things, this is what I will try to stick to from now on:

So if you write papers, maybe consider:

• Saying long-range or non-nearest-neighbour when the qubits sit on the same device.
• Saying distributed, remote or inter-QPU when they don't.
• If you do use "non-local", define it in one sentence. It costs you nothing and saves me a whole afternoon.

And, yes, I am aware the DQC folk are not blameless either... "Non-local" in physics already meant something (Bell non-locality)... So maybe we give it back?

AI warning