Complete Optimal Control Gate Sets

• Complete optimal control gate sets are families of quantum gate operations enabling universal synthesis while minimizing errors, calibration, and operational time.
• They integrate circuit synthesis with control-theoretic approaches to achieve native SU(4) coverage and optimized two-qubit interactions in various hardware platforms.
• Recent studies employ advanced objective functionals and continuous control formulations to enhance robustness, mitigate crosstalk, and achieve time-optimal quantum operations.

Complete optimal control gate sets are gate families or control resources that combine universality with resource optimality under a specified hardware model. In the broad quantum-information sense, a complete or universal gate set is any finite set of gates from which arbitrary unitary transformations on $n$ qubits can be synthesized; operationally, arbitrary single-qubit rotations together with any entangling two-qubit gate suffice. In a stricter control-theoretic sense, completeness means that the available drift and control Hamiltonians generate the full Lie group $SU(2^n)$ on the computational subspace. Recent work has used the same phrase for several related but distinct objectives: native realization of all of $SU(4)$ on two qubits without switching interaction mechanisms, device-wide single-qubit basis coverage coordinated against crosstalk, and continuous parameterized control surfaces that cover one- and two-qubit gate manifolds (Chen et al., 5 Feb 2025, Kairys et al., 2021, Goldschmidt et al., 16 Mar 2026).

1. Completeness and optimality as distinct design criteria

A complete gate set is not necessarily optimal. In the standard circuit model, completeness refers to synthesis capability: arbitrary single-qubit rotations plus any entangling two-qubit gate generate a universal set, and in controllability language a platform is universal if the available controls generate the full Lie group $SU(2^n)$ on the computational subspace (Kairys et al., 2021). By contrast, the two-qubit control literature has sharpened the meaning of optimality: for two-qubit operations, a complete and optimal control gate set is one that is complete in the sense of generating the entire special unitary group $SU(4)$ natively, without switching among multiple physical interaction mechanisms or addressing different transitions, and optimal in the sense of minimizing fundamental costs under the available Hamiltonian, notably entangling-gate count for universal synthesis and time-to-target for a given interaction strength and bounded controls (Chen et al., 5 Feb 2025).

This distinction matters because a compiled universal basis can be complete while still being suboptimal in calibration burden, leakage behavior, or execution time. The superconducting transmon literature makes this contrast explicit: traditional native gate sets often switch between distinct entangling mechanisms, such as using the $|11\rangle \leftrightarrow |20\rangle$ avoided crossing for CZ/CNOT-class gates and $|10\rangle \leftrightarrow |01\rangle$ exchange for iSWAP-family gates, whereas unified exchange-plus-drive control seeks full $SU(4)$ coverage from a single interaction mechanism (Chen et al., 5 Feb 2025).

A second distinction concerns scope. In the crosstalk-robust gate-set literature, completeness is defined at the device level rather than at the level of native arbitrary two-qubit synthesis. On IBM Eagle with heavy-hex connectivity, a complete gate set consists of $I$, $\sqrt{X}$, $SU(2^n)$0, and virtual $SU(2^n)$1, replicated in two colors sufficient for device-wide vertex coloring and coordinated so that parallel operation remains robust anywhere on the chip (Goldschmidt et al., 16 Mar 2026). This usage is compatible with native two-qubit entanglers and virtual phase updates, but it is not identical to native reachability of all of $SU(2^n)$2.

2. Controllability, local equivalence, and nonlocal geometry

For two qubits, the nonlocal content of a gate is naturally organized by the KAK, or Cartan, decomposition. Any $SU(2^n)$3 can be written as

$SU(2^n)$4

with

$SU(2^n)$5

where $SU(2^n)$6 lies in the Weyl chamber, a tetrahedron whose vertices are $SU(2^n)$7, $SU(2^n)$8, $SU(2^n)$9, and $SU(4)$0. Two unitaries are locally equivalent if and only if they share the same $SU(4)$1 (Chen et al., 5 Feb 2025).

This geometry separates mere universality from stronger notions of native expressivity. Any entangling two-qubit gate, combined with arbitrary single-qubit gates, yields a universal gate set. Perfect entanglers form a distinguished subset of $SU(4)$2: they are exactly those gates capable of creating a maximally entangled state from at least one product input state. In Weyl-chamber coordinates, the perfect-entangler region is the convex polyhedron defined by

$SU(4)$3

and can be characterized by local invariants $SU(4)$4. The functional

$SU(4)$5

vanishes on the perfect-entangler boundary, and the piecewise functional $SU(4)$6 introduced for optimal control satisfies $SU(4)$7 if and only if $SU(4)$8 is a perfect entangler (Watts et al., 2014). This is useful when the control objective is not a single canonical gate but any member of a desirable nonlocal class.

Within this geometry, the $SU(4)$9 gate occupies a special role. In canonical coordinates it sits at

$SU(2^n)$0

It is unique up to local equivalence in enabling exact synthesis of an arbitrary $SU(2^n)$1 unitary using only two applications of $SU(2^n)$2 plus local gates, whereas CNOT or iSWAP require three (Chen et al., 5 Feb 2025). This establishes an entangler-count notion of optimality that is stronger than ordinary universality.

3. Optimal-control formulations and target functionals

Optimal-control gate-set design is shaped by the choice of objective functional. A standard formulation chooses controls $SU(2^n)$3 for a time-dependent Hamiltonian $SU(2^n)$4 so that

$SU(2^n)$5

matches a target gate $SU(2^n)$6 up to an allowed equivalence (Kairys et al., 2021). The baseline objective uses

$SU(2^n)$7

which is global-phase invariant but penalizes locally equivalent two-qubit entanglers unequally. This is precisely the situation in which more structured objectives become useful.

Two such objectives were introduced to encode local equivalence and echo structure explicitly. The first,

$SU(2^n)$8

rewards any pulse whose two-qubit action can be converted to the target by pre- and post-single-qubit rotations. The second,

$SU(2^n)$9

targets single-pass controls that become the desired entangler only when echoed and interleaved with local rotations. In simulation on a three-level-per-transmon microwave-only model, these objectives produced simpler pulses, higher fidelities, and faster convergence than direct infidelity minimization $SU(4)$0 (Kairys et al., 2021).

A related line of work has generalized the target from a single gate to an entire gate manifold. In continuously parameterized settings, the goal is to learn a map $SU(4)$1 that implements $SU(4)$2 across a parameter domain. One direct-collocation template minimizes

$SU(4)$3

subject to discretized dynamics and bounds, while a coordinated “direct-sum” problem solves multiple parameter samples jointly with fidelity constraints $SU(4)$4 (Bhattacharyya et al., 2024). In that framework, pretraining a physics-informed neural network on direct-sum solutions reduced the number of epochs needed to reach $SU(4)$5 average fidelity from $SU(4)$6 to $SU(4)$7 for $SU(4)$8, and from $SU(4)$9 to $|11\rangle \leftrightarrow |20\rangle$0 for $|11\rangle \leftrightarrow |20\rangle$1. The same study also showed that transfer learning could recover average gate fidelity from $|11\rangle \leftrightarrow |20\rangle$2–$|11\rangle \leftrightarrow |20\rangle$3 back to $|11\rangle \leftrightarrow |20\rangle$4 for $|11\rangle \leftrightarrow |20\rangle$5 with fewer than $|11\rangle \leftrightarrow |20\rangle$6 calibration points, and to $|11\rangle \leftrightarrow |20\rangle$7 for $|11\rangle \leftrightarrow |20\rangle$8 with fewer than $|11\rangle \leftrightarrow |20\rangle$9 points (Bhattacharyya et al., 2024).

These formulations illustrate a recurring principle: optimal-control gate-set design increasingly targets equivalence classes, reachable regions, or continuous manifolds, rather than a single fixed matrix representative. That shift is central to the notion of completeness in realistic hardware settings.

4. Native and coordinated realizations on superconducting hardware

The most explicit experimental realization of a complete optimal control gate set for two qubits is the unified exchange-plus-drive scheme on a transmon–coupler–transmon unit cell. With both qubits tuned into resonance, $|10\rangle \leftrightarrow |01\rangle$0, the effective Hamiltonian is

$|10\rangle \leftrightarrow |01\rangle$1

where $|10\rangle \leftrightarrow |01\rangle$2. The operator set $|10\rangle \leftrightarrow |01\rangle$3 generates the full Lie algebra $|10\rangle \leftrightarrow |01\rangle$4, so the controls are fully two-qubit controllable. Experimentally, arbitrary Weyl-chamber points were realized natively without invoking $|10\rangle \leftrightarrow |01\rangle$5, and a broad suite of representative two-qubit unitaries achieved an average XEB fidelity of $|10\rangle \leftrightarrow |01\rangle$6. Representative values at $|10\rangle \leftrightarrow |01\rangle$7 MHz include $|10\rangle \leftrightarrow |01\rangle$8 at $|10\rangle \leftrightarrow |01\rangle$9 ns with error $SU(4)$0, iSWAP at $SU(4)$1 ns with error $SU(4)$2, a CNOT-class gate at $SU(4)$3 ns with error $SU(4)$4, SWAP at $SU(4)$5 ns with error $SU(4)$6, and the $SU(4)$7 gate at $SU(4)$8 ns with error $SU(4)$9. A calibrated $I$0 gate attained an XEB error of $I$1, and two-$I$2 synthesis reconstructed arbitrary Weyl-chamber points with an average error per $I$3 of $I$4 and a relative spread of approximately $I$5. The same control family reduced circuit depth for entangled-state preparation: a $I$6-qubit $I$7 state prepared with nine two-qubit gates reached $I$8 fidelity, and a $I$9-qubit double-excitation Dicke state $\sqrt{X}$0 synthesized with eight $\sqrt{X}$1 operations reached $\sqrt{X}$2, compared with a cited exact CNOT-based construction using $\sqrt{X}$3 CNOTs (Chen et al., 5 Feb 2025).

A different superconducting realization addresses completeness at device scale rather than at the level of native arbitrary $\sqrt{X}$4. Crosstalk-Robust Gate Sets are a complete, colored single-qubit basis for heavy-hex hardware, containing $\sqrt{X}$5, $\sqrt{X}$6, $\sqrt{X}$7, and virtual $\sqrt{X}$8, with two colors sufficient for device-wide vertex coloring. The coordinated optimal-control problem minimizes a sum of pairwise crosstalk susceptibilities over graph edges while enforcing $\sqrt{X}$9 gate fidelity constraints. On IBM Eagle, this device-wide gate set eliminated the fitted ZZ oscillation in XY4 dynamical decoupling and reduced the eight-qubit decay rate to $SU(2^n)$00 MHz, compared with $SU(2^n)$01 MHz for Gaussian gates and $SU(2^n)$02 MHz for default gates. In simultaneous four-qubit random Clifford circuits, it achieved $SU(2^n)$03 versus $SU(2^n)$04 for Gaussian gates, and the abstract reported a $SU(2^n)$05 ($SU(2^n)$06) improvement in gate fidelity for simultaneous four-qubit random Clifford circuits. In a four-qubit TFIM Hamiltonian simulation, CRGS and detuning-robust controls were approximately $SU(2^n)$07 more accurate after a single Trotter step than Gaussian gates and delivered a median approximately $SU(2^n)$08 algorithmic improvement across repetitions. Pulse-level simulations further indicated that stronger coupling can be exploited when the single-qubit gate set is optimized: at $SU(2^n)$09, CRGS yielded approximately $SU(2^n)$10 higher fidelity than detuning-robust controls and approximately $SU(2^n)$11 higher than Gaussian controls, supporting the statement that optimized gate sets can enable larger qubit–qubit coupling strengths that cut two-qubit gate times in half (Goldschmidt et al., 16 Mar 2026).

Earlier superconducting work already exhibited many ingredients now associated with complete optimal-control gate sets. For inductively coupled flux qubits at the sweet spot, Krotov optimization produced single-qubit $SU(2^n)$12 and $SU(2^n)$13 gates in $SU(2^n)$14–$SU(2^n)$15 ns and CNOT in approximately $SU(2^n)$16 ns, with decoherence-including errors $SU(2^n)$17–$SU(2^n)$18 for single-qubit gates and $SU(2^n)$19 for CNOT (Huang et al., 2014). Heavy-fluxonium control with automatic differentiation realized $SU(2^n)$20, $SU(2^n)$21, and $SU(2^n)$22 in $SU(2^n)$23 ns with closed-system fidelities $SU(2^n)$24, $SU(2^n)$25, and $SU(2^n)$26, and a $SU(2^n)$27 ns CZ with closed-system fidelity $SU(2^n)$28 and open-system fidelity $SU(2^n)$29 (Abdelhafez et al., 2019). Neighboring optimal control applied to a universal TRP gate set reduced one-qubit gate error probabilities to approximately $SU(2^n)$30 and a two-qubit modified controlled-phase gate to $SU(2^n)$31, below a target threshold of $SU(2^n)$32 (Peng et al., 2014).

5. Continuous, multilevel, and multi-qubit extensions

Complete optimal control gate sets are not restricted to discrete qubit libraries. One major extension is the continuous gate-set viewpoint, in which a parameter-indexed family $SU(2^n)$33 is synthesized directly rather than decomposed into a small fixed basis. In a superconducting model with local $SU(2^n)$34 control and exchange coupling, a physics-informed neural-network framework learned control surfaces for $SU(2^n)$35, $SU(2^n)$36, $SU(2^n)$37, and the entangling family

$SU(2^n)$38

Across $SU(2^n)$39 random test points, the learned $SU(2^n)$40 surface reached mean infidelity $SU(2^n)$41 with pretraining or $SU(2^n)$42 without it, and direct pulse-level implementation of $SU(2^n)$43 achieved comparable fidelity in approximately $SU(2^n)$44 ns rather than the $SU(2^n)$45–$SU(2^n)$46 ns required by a standard-basis decomposition into $SU(2^n)$47 CNOTs and $SU(2^n)$48–$SU(2^n)$49 single-qubit $SU(2^n)$50 rotations, a $SU(2^n)$51–$SU(2^n)$52 reduction in pulse-schedule duration (Bhattacharyya et al., 2024).

A related meta-optimization program treated continuous gate sets as parameter-conditioned pulse families. There, the learned single-qubit families

$SU(2^n)$53

were combined with direct CNOT synthesis and a continuous entangling family $SU(2^n)$54. For $SU(2^n)$55, the average infidelity was below $SU(2^n)$56 across the tested ranges, and for two-qubit $SU(2^n)$57 the learned controls maintained on-average infidelities $SU(2^n)$58–$SU(2^n)$59. The same work reported that a discrete CNOT-based decomposition of $SU(2^n)$60 can be approximately $SU(2^n)$61 slower than a direct continuous implementation (Preti et al., 2022).

A second extension concerns multilevel systems. In ultracold $SU(2^n)$62Rb, complete single-qutrit control in $SU(2^n)$63 was demonstrated using only two resonant microwave tones. The primitive single-tone rotations $SU(2^n)$64 and $SU(2^n)$65, together with a virtually implemented diagonal phase gate $SU(2^n)$66, form one complete gate set; a second replaces one single-tone pulse by the dual-tone synthesized coupling

$SU(2^n)$67

which effects a direct $SU(2^n)$68 rotation in a $SU(2^n)$69-scheme at

$SU(2^n)$70

Using both constructions, the Walsh–Hadamard Fourier transform was implemented with similar final-state fidelity and purity (2208.00045).

A third extension elevates the primitive itself from two-body to multi-qubit control. On neutral-atom Rydberg hardware, Krotov optimization produced one-step multi-qubit controlled-phase gates, controlled-rotation implementations of controlled-NOT family gates, and a three-qubit Fredkin gate. The amplitude-plus-phase-shaped Fredkin pulse reached $SU(2^n)$71 fidelity while accounting for spontaneous emission, laser fluctuations, and Doppler dephasing. Its average time-integrated Rydberg population was $SU(2^n)$72, compared with $SU(2^n)$73 for an $SU(2^n)$74-CNOT decomposition cited in the same study. For the optimized $SU(2^n)$75-$SU(2^n)$76 gate, $SU(2^n)$77, implying spontaneous-emission error of approximately $SU(2^n)$78 from $SU(2^n)$79 (Abedi et al., 27 Nov 2025). These results broaden the meaning of completeness from a universal two-local basis to a hardware-native set that already contains useful multi-qubit primitives.

6. Robustness, asymptotic optimality, and unresolved questions

The literature uses several inequivalent optimality criteria. In analog and pulse-level control, optimality often means minimum time, minimum entangler count, minimum sensitivity to crosstalk or detuning, or minimum calibration complexity. In fault-tolerant synthesis, by contrast, optimality is usually measured by non-Clifford resource count. For most multi-qubit controlled $SU(2^n)$80 operations in the Clifford+$SU(2^n)$81 model, the best known upper bound is

$SU(2^n)$82

with high probability, and this matches the lower bound

$SU(2^n)$83

when almost-controlled approximations are prohibited. The same framework gives ancilla-free $SU(2^n)$84 synthesis with

$SU(2^n)$85

improving over a cited KAK-based scaling of $SU(2^n)$86 (Yamazaki et al., 15 Mar 2026). This is an optimal-control gate-set result in a different sense from time-optimal analog synthesis, but it addresses the same question: what constitutes a complete set of control primitives at minimal cost?

Robustness has likewise been formalized beyond ad hoc pulse tuning. A recent geometric framework defines a chance-constrained robust optimal control problem in which the Hamiltonian undergoes stochastic switching,

$SU(2^n)$87

and seeks controls minimizing a risk functional $SU(2^n)$88 subject to

$SU(2^n)$89

Within that framework, robust completeness means that density of the reachable set must persist under bounded perturbations or stochastic switching, while robust gate complexity is defined by the minimal sequence length meeting an accuracy target with high probability (Aspman et al., 2024). This provides a common language for comparing device-level crosstalk-robust sets, offset-charge-randomized pulse families, and fault-tolerant discrete synthesis.

Two recurring misconceptions are clarified by this body of work. First, universality alone does not imply native maximal expressivity: a single perfect entangler plus local gates is universal, yet it does not mean that the platform natively reaches all of $SU(2^n)$90 in one entangling pulse. Second, native completeness at one layer does not guarantee completeness at another: a device-wide single-qubit basis such as CRGS is complete for parallel one-qubit moments on its topology, but it remains compatible with rather than equivalent to native arbitrary two-qubit control (Chen et al., 5 Feb 2025, Goldschmidt et al., 16 Mar 2026).

Several open directions remain explicit in the literature. Unified exchange-plus-drive control identifies formal global proofs of time-optimality under realistic amplitude constraints and scaling to multi-qubit networks with concurrent resonant blocks as open problems (Chen et al., 5 Feb 2025). Crosstalk-robust gate-set design identifies robust two-qubit envelopes, DRAG-like cancellation tones, reduced parametric control families, and improved crosstalk metrics as next steps (Goldschmidt et al., 16 Mar 2026). Continuous gate-surface methods point to region-aware interpolation for fractured minimum-time manifolds and to memory limits of direct-sum coordination as system size grows (Bhattacharyya et al., 2024). Together, these issues indicate that complete optimal control gate sets are best understood not as a single fixed object, but as a family of architecture-specific solutions balancing controllability, calibration, robustness, and the relevant cost metric for the computational layer of interest.