Superconducting Gates and Readout

A transmon sitting at 15 millikelvin inside a dilution refrigerator has two energy levels, |0\rangle and |1\rangle, separated by roughly 5 GHz. It can live there, undisturbed, for about 300 microseconds before decoherence eats the information. That is pretty long. But if nobody ever does anything to the transmon, you do not have a computer — you have a very cold, very expensive trinket.

You need three things. You need to rotate the qubit — flip |0\rangle to |1\rangle, or put it into a superposition. You need to entangle pairs of qubits. And you need to read the qubit out when you are done. All three happen over coaxial cables running into the fridge, all three use microwaves, and all three have to be done fast enough that you can fit thousands of gates into one coherence time.

This chapter is the manual. It covers single-qubit microwave drives, the three leading two-qubit-gate technologies, and dispersive readout. By the end you will know why a CNOT on IBM's hardware is actually a cross-resonance pulse, why Google's Willow chip uses tunable couplers instead, and why "we measured the qubit" actually means "we probed a resonator for 800 nanoseconds."

Single-qubit gates — a microwave pulse is a Bloch-sphere rotation

Start with the simplest thing. You have one transmon. You want to flip it from |0\rangle to |1\rangle.

The trick is resonance. The transmon has a specific transition frequency \omega_{10} — the energy gap between |0\rangle and |1\rangle, divided by \hbar. If you shine microwaves at the transmon at exactly that frequency, the electromagnetic field pushes the qubit to absorb a photon and climb to |1\rangle. Shine them longer, and the qubit oscillates between |0\rangle and |1\rangle — these are Rabi oscillations, the workhorse of every atomic and superconducting qubit experiment.

The drive Hamiltonian and the rotating-wave trick

A microwave drive on a transmon adds a time-dependent term to the Hamiltonian. In the qubit's two-level approximation, you can write the total Hamiltonian as

where Z and X are the Pauli operators on the qubit, \omega_d is the drive frequency, \varphi is the drive phase, and \Omega(t) is the pulse envelope (amplitude in rad/s).

Why this form: the first term is the qubit's own energy — its resonance at \omega_{10}. The second is the microwave pushing the charge on the transmon's superconducting island back and forth, which in the two-level truncation looks like an X operator multiplied by the instantaneous field amplitude. The phase \varphi controls which axis in the equatorial plane the pulse pushes along.

Go into a frame rotating at \omega_d. After the rotating-wave approximation (drop the terms oscillating at 2\omega_d — they average to nothing), the effective Hamiltonian becomes

with \Delta = \omega_{10} - \omega_d the detuning between qubit and drive.

Why the detuning matters: if \Delta = 0 (drive exactly on resonance with the qubit), the Z term vanishes and the drive implements a pure rotation in the equatorial plane — around +X if \varphi = 0, around +Y if \varphi = \pi/2. Any detuning twists the axis off the equator and reduces gate fidelity. This is why calibration obsesses over qubit-frequency drift.

On resonance, a pulse of integrated amplitude \theta = \int \Omega(t)\, dt applied at phase \varphi rotates the qubit by angle \theta around the axis \hat{n} = (\cos\varphi, \sin\varphi, 0). This gives you any rotation in the X-Y plane. Combined with virtual-Z (see below), you get every single-qubit unitary you want.

Pulse shapes — why Gaussian, why DRAG

A square pulse (on, constant amplitude, off) would work in principle but spreads energy across many frequencies — some of which would excite the transmon out of the qubit subspace into |2\rangle. That failure mode is called leakage, and it is the biggest enemy of fast superconducting gates. Transmons have low anharmonicity: the |1\rangle \to |2\rangle transition is only \sim 300 MHz below \omega_{10}, so a square pulse with significant spectral content at that distance will populate |2\rangle.

The fix is a smooth envelope. A Gaussian pulse of width \sigma has spectral width \sim 1/\sigma, so if \sigma \gtrsim 5 ns the spectral content at the |1\rangle \to |2\rangle transition is exponentially suppressed. But Gaussians alone leave a small phase error. DRAG — Derivative Removal by Adiabatic Gate (Motzoi et al. 2009) — adds a small amount of the pulse's time derivative on the other quadrature (the Y channel for an X drive, or vice versa). The DRAG coefficient is calibrated to cancel the leakage almost exactly. Modern hardware uses DRAG as a matter of course; the residual leakage is below 10^{-4} per gate.

The free R_z — virtual-Z gates

You might think R_z(\theta) rotations require a physical pulse too. They do not. Because R_z(\theta) = e^{-i\theta Z/2} commutes with the frame of measurement, you can implement it for free by redefining the phase of every subsequent pulse. If the next gate is R_x(\phi) = e^{-i\phi X/2}, you instead apply R_x at drive phase \varphi - \theta — which acts as if an R_z(\theta) happened just before.

This is virtual-Z (McKay et al. 2017). It is literally zero duration, zero error, and zero cost. It is the reason decompositions like H = R_z(\pi/2)\, R_x(\pi/2)\, R_z(\pi/2) are cheap on superconducting hardware — only the two R_x are real pulses; the two R_z are bookkeeping in the control software.

Net result: a transmon's native gate set is typically \{R_x(\pi/2), R_z(\varphi\,\text{any})\} for single qubits, plus one two-qubit entangler. Compilers translate arbitrary unitaries into this set.

Two-qubit gates — three ways, three companies

Single-qubit gates are straightforward. The hard part of a superconducting quantum computer is getting two qubits to interact in a controlled way. You need an entangling gate — something like a CNOT, CZ, or iSWAP — and you need it fast and accurate. Three architectures dominate the field in 2026.

Cross-resonance (IBM)

IBM builds fixed-frequency transmons, each detuned from its neighbours by a few hundred MHz. To entangle qubit A with qubit B, you apply a microwave drive to qubit A, at qubit B's frequency. The drive is off-resonant with A (so it barely excites A) and resonant with B (but B is not directly wired to the drive — A is). Through the shared bus resonator, the drive couples to B with an effective strength that depends on whether A is in |0\rangle or |1\rangle.

In the rotating frame, the effective interaction takes the form

where J_{ZX} is the cross-resonance coupling, proportional to the drive amplitude and to the inverse detuning between the qubits. Turning this on for time t = \pi / (2 J_{ZX}) implements \exp(-i\tfrac{\pi}{4} Z_A X_B) — a ZX rotation. With single-qubit wrappers, this is equivalent to a CNOT.

Why this architecture: no tunable elements means fewer wires, better manufacturability, and more stable frequencies. Cost: gate times are longer (200–400 ns) because J_{ZX} is weak, and spectator qubits can leak in crosstalk. IBM's engineering over 2017–2024 brought cross-resonance fidelity from 95% to 99.7% through echoed-cross-resonance sequences, careful calibration of static ZZ coupling, and custom DRAG-like pulse shaping.

IBM Heron (2024): cross-resonance CNOT fidelity 99.7% average, gate time ~300 ns.

iSWAP and tunable flux (Google's early path)

Google's Sycamore (2019, the chip that ran the famous 53-qubit random-circuit-sampling experiment) used a different recipe. Each transmon had a flux-biased SQUID — a DC-flux-tunable element — so its frequency could be shifted by several hundred MHz on a few-nanosecond timescale. To do a two-qubit gate, you tune two neighbouring qubits briefly into resonance. Their |01\rangle and |10\rangle states hybridize, swap populations, and pick up a phase — this is the iSWAP family.

Gate time: as short as 12 ns. Fidelity: 99.4% in 2019, improved since. Cost: extra flux-bias wires per qubit, sensitivity to flux noise (limits T_2 at the tuning sweet spot).

Parametric CZ via tunable couplers (Google Willow, Rigetti)

Modern Google hardware (Willow, 2024) and Rigetti's Ankaa keep the two data transmons fixed-frequency but insert a tunable coupler — a third transmon whose frequency can be modulated by AC flux. Modulating the coupler at the detuning frequency |\omega_A - \omega_B| drives a parametric exchange interaction. With the right pulse shape, the interaction implements a controlled-Z in 40–80 ns, and the coupler can be tuned to zero effective coupling when idle — killing spectator crosstalk.

Why this wins: fast (tens of ns), cleanly turnable on and off, lower crosstalk. Current state of the art: Google Willow 99.7% CZ, 40 ns; Rigetti 99.5%, 72 ns; IBM's newer Egret/Flamingo architectures are also adopting tunable couplers.

Readout — dispersive measurement

A transmon's state cannot be read directly by scoping the qubit wire — the microwave signal you would need to send would smash the state. Instead, you pair every qubit with a readout resonator (a small superconducting LC oscillator at a different frequency, say 7 GHz) and use a beautiful bit of physics called dispersive coupling.

The physics

The qubit and resonator interact via the Jaynes–Cummings Hamiltonian. In the dispersive regime — where the detuning \Delta = \omega_q - \omega_r is large compared to the bare coupling g — the interaction reduces to

where \chi = g^2 / \Delta is the dispersive shift (typically 0.5–5 MHz) and a^\dagger a is the photon-number operator of the resonator. The effect: the resonator's frequency depends on the qubit state. If the qubit is |0\rangle, the resonator sits at \omega_r + \chi. If |1\rangle, at \omega_r - \chi.

Why this is wonderful: sending a weak microwave probe tone at \omega_r into the resonator and measuring the phase of what comes back tells you the qubit state without, in principle, absorbing any energy from the qubit itself. The interaction is a pure Z-type coupling — it commutes with the qubit Hamiltonian — so the measurement is quantum non-demolition (QND) up to higher-order corrections. You can measure repeatedly and get the same answer.

The measurement protocol

1. Send a probe tone at frequency \omega_r (midway between the two qubit-conditioned peaks) for a readout window of t_m.
2. The reflected microwave is routed through a cold circulator and amplified by a Josephson parametric amplifier (JPA) or traveling-wave parametric amplifier (TWPA) at the 20 mK stage — these add quantum-limited noise. Then further amplification at 4 K (HEMT) and room temperature.
3. At room temperature, demodulate the signal into in-phase (I) and quadrature (Q) components.
4. Repeat the circuit many times. Plot each shot as a point on the (I, Q) plane. You see two blobs — one for each qubit state — separated by a distance set by the integrated signal-to-noise ratio (SNR).
5. Fit a threshold line. Each new shot is classified as |0\rangle or |1\rangle based on which side of the threshold its (I, Q) falls.

The readout fidelity is F = 1 - \tfrac{1}{2}(P(1|0) + P(0|1)) — the average probability of correct assignment.

Numbers

• Measurement time: 100 ns (fast, slightly lower fidelity) to 1 μs (slower, higher fidelity).
• Fidelity: 97–99.5% on 2024 hardware.
• Quantum-limited amplifiers: JPA (bandwidth ~20 MHz per qubit) or broadband TWPAs (2 GHz, tens of qubits simultaneously).
• Back-action: ideally just measurement-induced dephasing; in practice, photons leaking to the qubit environment produce a small T_1 degradation (\sim Purcell effect).

Error sources and fidelity budgets

Every gate has a fidelity budget. Understanding where the error comes from tells you what to optimise.

• Coherent errors (over-rotation, miscalibration). Pulse amplitude calibration drift gives R_x(\pi + \epsilon) instead of R_x(\pi). These are fixable by recalibration; modern hardware runs auto-calibration every few minutes.
• Incoherent errors (T_1, T_2). During a 30-ns gate with T_1 = 300\,\mus, the probability of a relaxation event is 30 / 3 \times 10^5 = 10^{-4}. This sets a fundamental floor on single-qubit-gate fidelity at about 1 - 10^{-4} = 99.99\%.
• Leakage to |2\rangle. With DRAG, typically below 10^{-4} per gate. Without DRAG, easily 10^{-2}.
• Crosstalk. Driving qubit A at \omega_B (cross-resonance) weakly drives any other qubit at frequency \omega_B — the "spectator" problem. Mitigated by echoes and careful frequency allocation.
• Readout back-action. Photons in the resonator cause a qubit frequency shift via the AC Stark effect, and dissipation through the Purcell effect. Modern designs use Purcell filters to suppress the latter.

A useful rule of thumb: on 2024 hardware, fidelity is limited roughly equally by the four listed sources. When one is improved, another becomes the bottleneck.

State of the art — 2024 to 2026

The field moves quickly. Benchmark numbers as of April 2026:

The trapped-ion column beats superconducting on every fidelity metric (by a factor of 3–10 on 2-qubit errors) and wins by orders of magnitude on coherence. The cost is speed: ion gates are 100–500 μs versus 30 ns single-qubit and 50 ns two-qubit for superconducting. A fair fidelity-per-unit-time comparison is more complicated; for now, both platforms are viable and invest in different bets on the road to fault tolerance.

Indian context

India's National Quantum Mission (2023, ₹6000 crore over 8 years) funds a superconducting-qubit effort led by TIFR Mumbai in collaboration with IISc Bangalore and IIT Bombay. The target is a 16-qubit processor by 2026, 50 qubits by 2028, 100 qubits by 2030. TIFR's group has published characterisation of two-qubit gates at the 99% level as of 2024 — the entry-level benchmark that IBM passed around 2017. The learning curve is steep but the direction is clear, and the hardware supply chain (dilution refrigerators from Bluefors in Finland, microwave electronics from Zurich Instruments) is established. For a student in India reading this chapter: the jobs will exist.

Common confusions

• "A gate is a pulse." Yes, specifically — on superconducting qubits, a gate is an amplitude-and-phase-shaped microwave pulse (or, for virtual-Z, a bookkeeping update). The abstract CNOT from a textbook is realised as a 300-ns cross-resonance sequence with three echo pulses embedded. The compiler translates the textbook gate into the hardware pulse.
• "Readout is the same as measurement." Close but not identical. Measurement is the quantum process — the projection that collapses the qubit. Readout is the specific engineering implementation: drive a resonator, measure the reflected microwave phase, threshold it. Some readouts are quantum non-demolition (repeatable); others destroy the qubit. The hardware matters.
• "Cross-resonance is the same as crosstalk." No. Crosstalk is an unwanted coupling between two qubits (a drive on A leaking onto B). Cross-resonance is a deliberately driven interaction between A and B, mediated by a shared bus resonator, where A is the controlled side and B is the target. The naming is unfortunate; in practice they are different phenomena, and the cross-resonance-gate designers spend a lot of effort to suppress actual crosstalk on neighbouring spectator qubits.
• "Two-qubit gates are fundamentally worse than single-qubit gates." Yes — by a factor of ~10 in error per gate, consistently across every platform. This is why circuit designers try to minimise two-qubit gate count. It is also why fault-tolerant thresholds are more stringent for two-qubit gates. The gap is slowly closing but not expected to vanish; two-qubit interactions are physically harder to turn on cleanly.
• "Virtual-Z is a trick that doesn't really work." It is exact — not an approximation. A virtual-Z is a change of the rotating frame in which the next pulse is applied. There is no physical error budget because there is no physical operation. The only subtlety is that virtual-Zs don't commute with measurement in the usual frame-tracking sense; the compiler handles this.
• "Higher photon number always gives better readout." Up to a point. More photons → more signal → better SNR. But photons in the resonator also cause the qubit frequency to shift via the AC Stark effect, and a high photon number eventually pushes the qubit into the non-dispersive regime where the QND property breaks down. Typical sweet spot: 3–10 photons.

Going deeper

DRAG — the derivation

The leakage problem is real. A Gaussian pulse of width \sigma has spectral density \propto e^{-\sigma^2 (\omega - \omega_{10})^2/2}. At the |1\rangle \to |2\rangle frequency \omega_{10} + \alpha (where \alpha \approx -300 MHz is the anharmonicity), the spectral weight is e^{-\sigma^2 \alpha^2 /2}. For \sigma = 5 ns and |\alpha| = 2\pi \times 300 MHz, this evaluates to e^{-0.5 \times 88} \approx 10^{-19} — absurdly small. So why does leakage happen at all?

Because the transmon is not a true two-level system. At leading order in perturbation theory, the drive on |0\rangle \to |1\rangle is accompanied by a next-order-in-\alpha off-resonant drive on |1\rangle \to |2\rangle, whose effect does not scale as the spectral density — it scales as 1/\alpha times the pulse amplitude. DRAG cancels this term exactly by adding a quadrature-phase pulse whose derivative matches the leakage-excitation amplitude. The original Motzoi et al. (2009) derivation gives

for a first-order cancellation. Higher-order DRAG schemes (Chen et al. 2016) cancel further leakage and phase errors and are standard on modern hardware.

Cross-resonance — the ZX derivation

Consider two transmons A (drive qubit) and B (target qubit) coupled via a shared bus resonator with exchange coupling J. The bare Hamiltonian (after RWA and adiabatic elimination of the resonator) is

Apply a drive \Omega \cos(\omega_B t)\, X_A. Go into a frame rotating at \omega_B for both qubits. Second-order perturbation theory in J/(\omega_A - \omega_B) produces, in the qubit-qubit effective Hamiltonian, the term

Over time t = \pi / (2 |H_{ZX}|) this implements \exp(-i\tfrac{\pi}{4} Z_A X_B). Single-qubit R_z(\pi/2) and R_y(\pi/2) wrappers convert this to CNOT. Modern IBM hardware uses an echoed cross-resonance sequence (drive, X_A, drive, X_A) to cancel the unwanted IX and ZI terms exactly — the ZX and IX split into even and odd sectors.

Dispersive readout — the shift formula

Start from the Jaynes–Cummings Hamiltonian for a qubit (two-level atom with frequency \omega_q) coupled to a single resonator mode (frequency \omega_r, bosonic operator a):

Apply the Schrieffer–Wolff transformation at small g/|\Delta| where \Delta = \omega_q - \omega_r:

with \chi = g^2 / \Delta. The \chi Z a^\dagger a term is the readout signal. For a transmon (which is weakly anharmonic, not a true two-level system) there is a correction \chi = g^2/\Delta \cdot \alpha / (\alpha + \Delta) that accounts for the |2\rangle level — this is the formula used in real calibration.

Two-level systems (TLS) — the coherence limit

Transmons do not sit in empty space. They live on an amorphous dielectric (usually aluminium oxide) and interact with defects in that dielectric that themselves behave as two-level systems — TLS defects. Each TLS couples to the qubit via the dipole-dipole interaction, providing a channel for T_1 decay. TLS defects are the dominant source of T_1 variation across qubits on the same chip. They appear and disappear as the chip is thermally cycled. The field has spent two decades trying to reduce TLS density by improving fabrication (tantalum-on-sapphire substrates give T_1 \sim 1 ms, aluminium-on-silicon gives \sim 100–300\,\mus); progress is incremental and active.

The fault-tolerance threshold

For a surface code, the per-gate error rate must be below the threshold of about 1\% for the code to improve logical error with increasing distance. Current hardware is at 0.3–0.5\% per two-qubit gate — a factor of 2–3 below threshold. This is enough for small logical qubits (Google Willow demonstrated below-threshold performance at distance 7 in 2024), but not enough for truly useful encoded computation, which requires thresholds 100–1000× below. That factor is the engineering gap between NISQ and fault-tolerant computing.

NVIDIA cuQuantum and classical simulators

Superconducting hardware is benchmarked against classical simulators running the same circuit. NVIDIA's cuQuantum library lets a single GPU simulate dense state vectors of up to ~32 qubits and tensor-network approximations of much larger circuits. Benchmarking a 99.7%-fidelity 2-qubit gate on hardware means nothing without a classical baseline of what the gate should have produced. Every gate-fidelity claim you read is validated against such a simulator.

The IBM-versus-Quantinuum debate

IBM argues superconducting wins in the long run because it scales faster (1000+ qubits already) and gate speed lets you fit deeper circuits in a coherence time. Quantinuum argues trapped ions win because 99.9% gate fidelity and second-long coherence mean logical qubits today, and QCCD architectures (ion shuttling across junctions) will scale. The academic consensus: both will succeed for different applications. NISQ-scale chemistry simulations fit comfortably on either. Large-scale factoring probably needs millions of physical qubits regardless of platform, which means the winning platform is whichever first demonstrates reliable fabrication at that scale.

Where this leads next

• Trapped Ions — Paul Traps — the competing platform, with details on why its gates are 100× slower but 10× more accurate.
• Superconducting Transmon — the qubit itself (previous chapter).
• Decoherence — T1 and T2 — what sets the coherence-time ceiling.
• Photonic Quantum Computing — the no-fridge alternative.
• Neutral Atom Arrays — another Rydberg-based platform.

References

1. Philip Krantz et al., A Quantum Engineer's Guide to Superconducting Qubits (2019), Applied Physics Reviews — arXiv:1904.06560.
2. Alexandre Blais et al., Circuit Quantum Electrodynamics (2021), Reviews of Modern Physics — arXiv:2005.12667.
3. Rami Barends et al., Superconducting quantum circuits at the surface code threshold for fault tolerance (2014), Nature — arXiv:1402.4848.
4. IBM Quantum, IBM Quantum Hardware Documentation — ibm.com/quantum/technology.
5. Wikipedia, Superconducting quantum computing.
6. John Preskill, Lecture Notes on Quantum Computation, Chapter 7 — theory.caltech.edu/~preskill/ph229.