One Gate Scheme to Rule Them All: Introducing a Complex Yet Reduced Instruction Set for Quantum Computing

Classical chips keep instruction sets small because every instruction costs silicon — quantum chips do not work that way, so this paper builds one control scheme whose knobs reach every possible two-qubit operation, each in the fastest time physics allows.

Why this problem is worth solving

A

The paper’s opening move is to notice that the reason classical architectures favour small instruction sets does not transfer. On a classical chip, every instruction is combinational logic that must be hardwired, so a richer instruction set means more silicon and more cost. On a quantum chip, the gates are the qubits — an operation is applied by shaping a control signal, not by adding a circuit. Two gates that share a control scheme can be realised by the same hardware even if they do different things. So instruction set complexity lands on the difficulty of controlling qubits, not on the price of the chip, and a rich instruction set becomes affordable in a way it never was classically. The question this opens is whether every two-qubit operation could be native, delivered by one scheme with different settings.

What people did before, and where it broke

• Pick one native two-qubit gate. The controlled-NOT is the conventional choice, and everything else is compiled into sequences of it. Simple to calibrate, and an arbitrary two-qubit gate costs up to three of them.
• Pick a different single gate. Swap-like gates are natural on some superconducting hardware and are cheaper for some circuits, but the same three-application worst case applies.
• Support a continuous family. Recent hardware exposes tunable gate families, so a range of operations becomes directly available. A genuine advance and still a subset of the possibilities.
• Repeat a calibrated gate to reach a family. One approach avoids calibrating a continuous set by applying a single calibrated gate several times to generate the family. That works, and applying a gate several times compromises time optimality.
• Treat unwanted coupling as a feature. Transversely coupled qubits suffer a stray interaction that is the main coherent error. Rather than fight it, some designs add the gates it produces to the instruction set — which works but yields gates that are awkward to analyse, sometimes requiring brute-force search to compile with.
• Engineer the stray coupling away. Level engineering can suppress it, at the cost of device design constraints, and for most superconducting devices it lingers anyway.
• The gap. Nobody had asked whether a single control scheme could reach every two-qubit operation, and nobody had a design where the stray coupling was neither an error to suppress nor a special case to bless, but simply another term the scheme solves around.

How one scheme reaches every two-qubit gate

The idea

Spend the resource that is cheap on this architecture. The framing that makes the title’s paradox resolve is an argument about cost models. Reduced instruction sets are a good idea classically because instructions are logic gates etched into a die, so more of them means a bigger, costlier chip. On a superconducting quantum device the situation inverts: the quantum operations are integrated into the qubits themselves, and what a richer instruction set costs is control complexity — more elaborate pulse shaping and more calibration — rather than area. Gates that share an underlying control scheme can be delivered by identical hardware with different parameter settings. So the design becomes “complex yet reduced”: complex in what it can express, reduced in that a single scheme with four knobs delivers all of it. Choosing to spend whichever resource a given architecture makes abundant, rather than importing a cost model from a different one, is the same reasoning that justifies dedicating a fifth of a quantum machine’s qubits to shared communication infrastructure, covered elsewhere on this site.

Absorb the nuisance term instead of fighting it or blessing it. Transversely coupled qubits suffer a stray interaction that dominates the coherent error budget. There are two established responses: engineer it away with device design, or declare it part of the instruction set and compile with whatever gates it happens to produce. The first constrains the hardware and rarely eliminates the problem; the second yields gates that are hard to reason about and may need brute-force search to use. This scheme takes a third route — put the stray term into the Hamiltonian being solved and adjust the four control parameters to compensate. The result proved is stronger than the original claim: the scheme spans every two-qubit gate with the stray coupling present, provided it is no larger than the intended coupling, which holds in most experimental settings. On one gate family the stray coupling even makes the optimal time shorter. Treating a device’s imperfection as a parameter the design solves around, rather than an error to suppress or a quirk to exploit ad hoc, is the same stance taken by work that navigates the noise landscape of a device rather than assuming it away, covered elsewhere on this site.

Calibrate the model, not the gates. The obvious objection to a continuous instruction set is that you cannot calibrate infinitely many gates. The paper answers at two levels. For a single gate, it observes that what matters is the three-number coordinate rather than the single-qubit components, since single-qubit gates are already high fidelity — and it constructs a way to extract those three numbers by turning the problem into phase estimation, using a doubled version of the gate that the scheme itself can produce by playing its waveforms in reverse with flipped phase and detuning. For the whole set, the argument is that because every gate shares one control scheme, the relationship between requested parameters and achieved fidelity should follow a simple mathematical model — so the task becomes fitting that model’s parameters, using randomised benchmarking over the gate family as the objective. Replacing enumeration over an enormous or infinite space with fitting a parameterised description you can optimise directly is the same manoeuvre behind making tensor program search differentiable, covered elsewhere on this site.

Following a request for an arbitrary two-qubit gate

Every two-qubit operation has exactly one such coordinate, so reaching all of them is the same as covering a bounded three-dimensional region. That equivalence is what turns "support every gate" into a solvable geometry problem.

Why it works

• Every two-qubit gate is three numbers. Stripping off single-qubit rotations leaves a coordinate in a bounded region → covering the region is equivalent to supporting every gate up to cheap corrections → so a scheme with enough tunable parameters can be shown to be complete rather than merely broad.
• Four knobs are enough to reach all of it. Two drive amplitudes, a detuning and a duration → the main theorem shows a setting exists for every coordinate → and at the optimal setting at least one of the three control parameters is zero, which simplifies what has to be produced.
• One gate replaces up to three. Any two-qubit operation is directly realisable → circuits need the theoretically minimal number of two-qubit instructions → which compounds with shorter individual gates rather than trading against them.
• Optimal time is attainable almost everywhere. For each target, parameters exist reaching the minimum time physics permits → averaged over random two-qubit gates the time is about 1.34 divided by the coupling strength → against 1.74 for the best comparable scheme and 6.66 for a flux-tuned controlled-Z.
• The stray coupling is compensated rather than suppressed. It enters the Hamiltonian the algorithm solves → the parameters returned already account for it → so the main coherent error source stops being an error at all, and for swap gates it even shortens the optimal time.

What it costs (trade-offs and scope)

• It has never been built. The authors say so plainly: the scheme has not been implemented in hardware. Everything here is analysis and numerical simulation, supported by the observation that a previous experiment used the same physical setup.
• Calibration is the standing objection and it is answered by proposal. The methods for calibrating a continuous family are described but not demonstrated, and one promising route — interpolating between calibrated parameter settings — is explicitly labelled possible but unproven.
• Optimal time is unreachable near the identity. Achieving it for gates close to doing nothing requires drive amplitudes growing without bound, so a cutoff is needed. The numerics show a practical cutoff costs almost nothing overall, but the guarantee is not uniform.
• Square pulses are assumed. Real pulses ramp up and down, which the analysis does not model. That imperfection also breaks a symmetry the calibration scheme relies on, which the paper acknowledges and works around.
• Device physics may push back. On transmon qubits leakage is the primary concern, requiring large anharmonicity and relatively small coupling — and small coupling means slower gates, which is exactly what the scheme is trying to avoid. Fluxonium avoids leakage but introduces phase errors from higher levels.
• The synthesis results are near, not at, the lower bound. Three-qubit gates need at most eleven generic two-qubit gates against a theoretical floor of six; the general n-qubit result sits within a factor of three of its bound.
• Compilers do not yet exist for this. Achieving the optimal instruction count in practice requires compilation schemes that can actually exploit an arbitrary two-qubit gate as the native instruction, which the authors list as future work.

The evaluation

Check yourself

Sources

The paper is “One Gate Scheme to Rule Them All: Introducing a Complex Yet Reduced Instruction Set for Quantum Computing” by Jianxin Chen, Dawei Ding, Weiyuan Gong, Cupjin Huang and Qi Ye, published at ASPLOS 2024 Volume 2, pages 779-796, DOI 10.1145/3620665.3640386. The authors are with Alibaba Quantum Laboratory and Tsinghua University, and the gate scheme is named AshN.

This explainer is written from the full text, read via the arXiv version at arxiv.org/abs/2312.05652. All figures above come from the paper’s own sections 4 through 7: the theorem that the scheme reaches every Weyl chamber coordinate at the optimal gate time for any coupling and any stray coupling no larger than it, the average random two-qubit gate time of approximately 1.341 divided by the coupling strength against 1.736, 4.712 and 6.664 for the compared schemes, the synthesis bound of at most eleven generic two-qubit gates for three-qubit unitaries against twenty controlled-NOT gates analytically with lower bounds of six and fourteen respectively, the general n-qubit result at three quarters of the quantum Shannon decomposition count, the quantum volume simulations in Cirq over 1350 random circuits per point with single-qubit error fixed at 0.1% and two-qubit error from 0.7% to 1.7%, and the observation that a practical amplitude cutoff performs almost identically to the unbounded optimum. The authors state in their introduction that the scheme has not yet been implemented in hardware, and in section 7 that interpolation of calibrated parameters is a possible but unproven route to reducing calibration overhead.