Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Two-Qubit Gates & Tunable Couplers

Cross-resonance, CZ via flux, and coupler designs.

One transmon is easy to control; two that talk to each other are where the errors live. Two-qubit gates are typically several times slower and an order of magnitude less accurate than single-qubit ones. This lesson derives the main mechanisms (static ZZZZ coupling, the cross-resonance interaction and the flux-activated CZ) and shows how tunable couplers switch interactions off when they are not wanted.

What you’ll learn

  • Derive the static ZZZZ shift of two coupled transmons from second-order perturbation theory.
  • Derive the cross-resonance ZXZX and IXIX rates for a transmon control qubit, and explain the echo.
  • Explain the CZ gate via the ∣11⟩\ket{11}–∣20⟩\ket{20} avoided crossing and compute its duration.
  • Analyze tunable couplers as interfering coupling paths, and discuss crosstalk, frequency collisions and the state of the art.

Coupled transmons and static ZZ

Two transmons coupled through a capacitor are well described, in the rotating-wave approximation, by

Hℏ=∑j=1,2(ωjbj†bj+αj2 bj†bj†bjbj)+g (b1†b2+b1b2†)\frac{H}{\hbar} = \sum_{j=1,2}\Big(\omega_jb_j^\dagger b_j+\frac{\alpha_j}{2}\,b_j^\dagger b_j^\dagger b_jb_j\Big)+g\,\big(b_1^\dagger b_2+b_1b_2^\dagger\big)
Each transmon is a Kerr oscillator with anharmonicity αj<0\alpha_j<0; g/2πg/2\pi is typically a few MHz for direct capacitive coupling.

When the qubits are detuned, ∣Δ∣≫g\abs\Delta\gg g with Δ=ω1−ω2\Delta = \omega_1-\omega_2, they do not exchange excitations, but the coupling still shifts the levels. The shift of ∣11⟩\ket{11} is not the sum of the shifts of ∣10⟩\ket{10} and ∣01⟩\ket{01}, because only ∣11⟩\ket{11} couples to the doubly excited states ∣20⟩\ket{20} and ∣02⟩\ket{02}:

DerivationStatic ZZ to second order
  1. δE10=g2Δ,δE01=−g2Δ\delta E_{10} = \frac{g^2}{\Delta},\qquad\delta E_{01} = -\frac{g^2}{\Delta}

    ∣10⟩\ket{10} and ∣01⟩\ket{01} repel each other through the matrix element gg.

  2. ⟨20|g b1†b2|11⟩=2 g,E11−E20=ω1+ω2−(2ω1+α1)=−(Δ+α1)\mel{20}{g\,b_1^\dagger b_2}{11} = \sqrt2\,g,\qquad E_{11}-E_{20} = \omega_1+\omega_2-(2\omega_1+\alpha_1) = -(\Delta+\alpha_1)

    The bosonic factor 2\sqrt2 appears because b1†∣1⟩=2∣2⟩b_1^\dagger\ket1 = \sqrt2\ket2.

  3. δE11=2g2−(Δ+α1)+2g2Δ−α2\delta E_{11} = \frac{2g^2}{-(\Delta+\alpha_1)}+\frac{2g^2}{\Delta-\alpha_2}

    Second-order shifts from ∣20⟩\ket{20} and, in the same way, from ∣02⟩\ket{02}, for which E11−E02=Δ−α2E_{11}-E_{02} = \Delta-\alpha_2.

  4. ζ:=E11−E10−E01+E00=2g2(α1+α2)(Δ+α1)(Δ−α2)\zeta := E_{11}-E_{10}-E_{01}+E_{00} = \frac{2g^2(\alpha_1+\alpha_2)}{(\Delta+\alpha_1)(\Delta-\alpha_2)}

    The single-excitation shifts cancel, leaving the ZZZZ rate: the frequency shift of one qubit conditioned on the other being excited. It vanishes for harmonic oscillators, αj=0\alpha_j = 0.

Check your understanding

Why does the static ZZZZ shift vanish for two coupled harmonic oscillators (α1=α2=0\alpha_1 = \alpha_2 = 0)?

  1. Because harmonic oscillators cannot be coupled
  2. Because the levels of coupled linear oscillators are sums of normal-mode energies, so E11−E10−E01+E00=0E_{11}-E_{10}-E_{01}+E_{00} = 0 exactly
  3. Because gg must vanish for harmonic oscillators
  4. Because the rotating-wave approximation fails
Show the answer

B. Because the levels of coupled linear oscillators are sums of normal-mode energies, so E11−E10−E01+E00=0E_{11}-E_{10}-E_{01}+E_{00} = 0 exactly

For linear oscillators the normal modes are exact and energies add, so there is no conditional shift; in the formula, α1+α2=0\alpha_1+\alpha_2 = 0 kills the numerator. Conditional dynamics, and hence entangling gates, need nonlinearity.

Cross-resonance: an all-microwave gate

Many processors use fixed-frequency transmons, which need no flux line and suffer no flux noise. To entangle them, drive the control qubit 1 at the frequency of the target qubit 2 (Rigetti & Devoret 2010; Chow et al. 2011). In the frame rotating at ωd=ω2\omega_d = \omega_2 the drive reads Ω2(b1+b1†)\frac{\Omega}{2}(b_1+b_1^\dagger), off-resonant from qubit 1 by Δ=ω1−ω2\Delta = \omega_1-\omega_2; call the exchange coupling JJ (the gg of the previous step).

DerivationEffective cross-resonance Hamiltonian
  1. ak=−ΩJ2 k+1Δ+kα1+ΩJ2 kΔ+(k−1)α1a_k = -\frac{\Omega J}{2}\,\frac{k+1}{\Delta+k\alpha_1}+\frac{\Omega J}{2}\,\frac{k}{\Delta+(k-1)\alpha_1}

    With qubit 1 in level ∣k⟩\ket k, the drive and the coupling combine at second order into an effective drive ak Xa_k\,X on qubit 2. One virtual path passes through ∣k+1⟩1\ket{k+1}_1 (energy denominator Δ+kα1\Delta+k\alpha_1), the other through ∣k−1⟩1∣1⟩2\ket{k-1}_1\ket1_2 (denominator Δ+(k−1)α1\Delta+(k-1)\alpha_1); the bosonic matrix elements contribute the factors k+1k+1 and kk.

  2. a0=−ΩJ2Δ,a1=ΩJ2Δ−ΩJΔ+α1a_0 = -\frac{\Omega J}{2\Delta},\qquad a_1 = \frac{\Omega J}{2\Delta}-\frac{\Omega J}{\Delta+\alpha_1}

    The target's drive depends on the control's state: the origin of the entanglement.

  3. Heff=a0+a12 I⊗X+a0−a12 Z⊗XH_{\text{eff}} = \frac{a_0+a_1}{2}\,I\otimes X+\frac{a_0-a_1}{2}\,Z\otimes X

    Write ∣0⟩ ⁣⟨0∣=(I+Z)/2\dyad0 = (I+Z)/2 and ∣1⟩ ⁣⟨1∣=(I−Z)/2\dyad1 = (I-Z)/2 on the control.

  4. Heff=ωZX2 Z⊗X+ωIX2 I⊗X,ωZX=−JΩΔ α1Δ+α1,ωIX=−JΩΔ+α1H_{\text{eff}} = \frac{\omega_{ZX}}{2}\,Z\otimes X+\frac{\omega_{IX}}{2}\,I\otimes X,\qquad\omega_{ZX} = -\frac{J\Omega}{\Delta}\,\frac{\alpha_1}{\Delta+\alpha_1},\qquad\omega_{IX} = -\frac{J\Omega}{\Delta+\alpha_1}

    To first order in Ω\Omega. In the two-level limit α1→∞\alpha_1\to\infty, ωZX→−JΩ/Δ\omega_{ZX}\to-J\Omega/\Delta and ωIX→0\omega_{IX}\to0. The transmon's second level both modifies the ZXZX rate and creates a spurious IXIX term (compare Magesan & Gambetta 2020).

A ZXZX rotation by π/2\pi/2, e−iπ4Z⊗Xe^{-i\frac\pi4Z\otimes X}, equals a CNOT up to single-qubit gates, so the gate time is t=π/(2∣ωZX∣)t = \pi/(2\abs{\omega_{ZX}}). The rate is small, ∝JΩ/Δ\propto J\Omega/\Delta, and the drive cannot be made arbitrarily strong without exciting the control, so cross-resonance gates typically take a few hundred ns. Note the pole at Δ=−α1\Delta = -\alpha_1: when the control's 1→21\to2 transition hits the target frequency the perturbative picture breaks down. It is one of several frequency collisions that fixed-frequency designs must avoid.

CZ via the |11⟩–|20⟩ avoided crossing

With frequency-tunable transmons, the strongest conditional interaction is the very one that produced ZZZZ. Tune qubit 1 with a flux pulse until ω1+ω2=2ω1+α1\omega_1+\omega_2 = 2\omega_1+\alpha_1, that is Δ=−α1\Delta = -\alpha_1, the pole of the ZZZZ formula. In the two-dimensional subspace {∣11⟩,∣20⟩}\{\ket{11},\ket{20}\} the coupling is 2g\sqrt2g, and no other computational state is resonant with anything (Strauch et al. 2003; DiCarlo et al. 2009).

DerivationA diabatic CZ
  1. H{11,20}=2 g σx(on resonance, rotating frame)H_{\{11,20\}} = \sqrt2\,g\,\sigma_x\qquad\text{(on resonance, rotating frame)}

    A two-level system with Rabi frequency 22g2\sqrt2g.

  2. e−i2g t σx∣t=π/(2g)=−Ie^{-i\sqrt2g\,t\,\sigma_x}\Big|_{t = \pi/(\sqrt2g)} = -\id

    After one full oscillation, ∣11⟩→∣20⟩→−∣11⟩\ket{11}\to\ket{20}\to-\ket{11}: the population returns, and ∣11⟩\ket{11} picks up a minus sign.

  3. ∣00⟩→∣00⟩,∣01⟩→eiϕ01∣01⟩,∣10⟩→eiϕ10∣10⟩,∣11⟩→−ei(ϕ01+ϕ10)∣11⟩\ket{00}\to\ket{00},\quad\ket{01}\to e^{i\phi_{01}}\ket{01},\quad\ket{10}\to e^{i\phi_{10}}\ket{10},\quad\ket{11}\to-e^{i(\phi_{01}+\phi_{10})}\ket{11}

    The other states only acquire single-qubit dynamical phases, removed with virtual ZZ gates. The conditional phase ϕ11−ϕ10−ϕ01+ϕ00=π\phi_{11}-\phi_{10}-\phi_{01}+\phi_{00} = \pi makes this a CZ.

Predict before you look

For g/2π=10g/2\pi = 10 MHz, how long does this diabatic CZ take, in ns? (t=π/(2g)t = \pi/(\sqrt2g), with gg in rad/s.)

Show the answer

35 ns

t=π/(2×2π×107 s−1)=1/(22×107)t = \pi/(\sqrt2\times2\pi\times10^7\ \text{s}^{-1}) = 1/(2\sqrt2\times10^7) s ≈35\approx35 ns. Practical CZ gates take a few tens of ns and use adiabatic or "fast-adiabatic" pulse shapes (Martinis & Geller 2014) to keep leakage into ∣20⟩\ket{20} low.

Tunable couplers, crosstalk and the frontier

Fixed coupling is a compromise: strong enough for fast gates means strong ZZZZ while idling. A tunable coupler is a third element, often itself a flux-tunable transmon with frequency ωc\omega_c above the qubits, coupled to each qubit (g1g_1, g2g_2), alongside a small direct coupling g12g_{12} between the qubits (Yan et al. 2018).

DerivationTwo interfering paths
  1. g~≈g12+g1g22(1Δ1+1Δ2),Δj=ωj−ωc\tilde g\approx g_{12}+\frac{g_1g_2}{2}\Big(\frac{1}{\Delta_1}+\frac{1}{\Delta_2}\Big),\qquad\Delta_j = \omega_j-\omega_c

    Eliminating the coupler in the dispersive regime, ∣Δj∣≫gj\abs{\Delta_j}\gg g_j, gives an effective exchange that sums the direct path and a virtual path through the coupler. Counter-rotating terms add corrections of the same form with ωj+ωc\omega_j+\omega_c in the denominators (Yan et al. 2018).

  2. ωc>ω1,2  ⇒  Δj<0:g~=0  when  g1g22(1∣Δ1∣+1∣Δ2∣)=g12\omega_c>\omega_{1,2}\;\Rightarrow\;\Delta_j<0:\qquad\tilde g = 0\ \text{ when }\ \frac{g_1g_2}{2}\Big(\frac{1}{\abs{\Delta_1}}+\frac{1}{\abs{\Delta_2}}\Big) = g_{12}

    With the coupler above the qubits the two paths have opposite signs, so one particular coupler frequency switches the interaction off; moving the coupler closer to the qubits turns it on.

  3. idle: g~≈0, ζ≈0;gate: g~ large ⇒ fast CZ or iSWAP-type gates\text{idle: }\tilde g\approx0,\ \zeta\approx0;\qquad\text{gate: }\tilde g\ \text{large}\ \Rightarrow\ \text{fast CZ or iSWAP-type gates}

    The residual ZZZZ is nulled at a nearby, not identical, coupler bias, because the coupler's own higher levels also contribute.

SchemeHardwareTypical durationMain issues
Echoed cross-resonancefixed-frequency qubits, fixed coupling~200–500 nsslow; frequency collisions; static ZZZZ; spectator errors
Flux-activated CZtunable qubits and/or couplers~20–60 nsflux noise away from sweet spots; leakage to ∣20⟩\ket{20}; flux-pulse distortions
Tunable-coupler CZ, iSWAP or fSimtunable coupler between the qubits~20–60 nsmore control lines; coupler calibration
Parametric and microwave-activated gatesmodulated couplers, or fixed qubits with strong drives~100–300 nsdrive-induced shifts; calibration
Orders of magnitude as of the mid-2020s. The best two-qubit errors are around 10−310^{-3} on good pairs, with medians across large chips several times higher.

Key takeaways

  • Coupled transmons acquire a static ZZZZ shift ζ=2g2(α1+α2)/[(Δ+α1)(Δ−α2)]\zeta = 2g^2(\alpha_1+\alpha_2)/[(\Delta+\alpha_1)(\Delta-\alpha_2)], a coherent error unless cancelled or switched off.
  • Cross-resonance makes the target's drive depend on the control's state: ωZX=−JΩΔα1Δ+α1\omega_{ZX} = -\frac{J\Omega}{\Delta}\frac{\alpha_1}{\Delta+\alpha_1}, plus a spurious ωIX=−JΩ/(Δ+α1)\omega_{IX} = -J\Omega/(\Delta+\alpha_1); the echo removes IXIX, ZIZI and ZZZZ to first order.
  • Flux-tuning to the ∣11⟩\ket{11}–∣20⟩\ket{20} resonance for one full oscillation gives a CZ in t=π/(2g)t = \pi/(\sqrt2g), about 35 ns for g/2π=10g/2\pi = 10 MHz; leakage is the main risk.
  • Tunable couplers create interfering coupling paths whose cancellation switches interactions off, suppressing ZZZZ and enabling fast gates; crosstalk and uniformity dominate at scale.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

The static ZZZZ shift ζ=2g2(α1+α2)(Δ+α1)(Δ−α2)\zeta = \frac{2g^2(\alpha_1+\alpha_2)}{(\Delta+\alpha_1)(\Delta-\alpha_2)} diverges at Δ=−α1\Delta = -\alpha_1. What happens physically there?

  1. ∣11⟩\ket{11} becomes resonant with ∣20⟩\ket{20}, perturbation theory breaks down and the two states hybridize strongly: the working point of a CZ gate
  2. The qubits decouple
  3. ∣10⟩\ket{10} becomes resonant with ∣01⟩\ket{01}
  4. Qubit 1 spontaneously leaks into its third level
Show the answer

A. ∣11⟩\ket{11} becomes resonant with ∣20⟩\ket{20}, perturbation theory breaks down and the two states hybridize strongly: the working point of a CZ gate

At Δ=−α1\Delta = -\alpha_1 the energies E11=ω1+ω2E_{11} = \omega_1+\omega_2 and E20=2ω1+α1E_{20} = 2\omega_1+\alpha_1 coincide, so second-order perturbation theory fails and the levels anticross with splitting 22g2\sqrt2g. Flux-activated CZ gates exploit exactly this point; fixed-frequency designs must avoid it.

Question 2

In the two-level limit, what is the leading cross-resonance ZXZX rate, and why are CR gates relatively slow?

  1. ωZX=Ω\omega_{ZX} = \Omega: limited only by drive power
  2. ωZX=J\omega_{ZX} = J: limited by the coupling
  3. ωZX≈−JΩ/Δ\omega_{ZX}\approx-J\Omega/\Delta: suppressed by the small ratio J/ΔJ/\Delta, while Ω\Omega cannot be large without exciting the control
  4. ωZX=Δ\omega_{ZX} = \Delta: limited by the detuning
Show the answer

C. ωZX≈−JΩ/Δ\omega_{ZX}\approx-J\Omega/\Delta: suppressed by the small ratio J/ΔJ/\Delta, while Ω\Omega cannot be large without exciting the control

The conditional drive is second order: the drive Ω\Omega times the dressing J/ΔJ/\Delta. With J/Δ∼0.03J/\Delta\sim0.03 and Ω/2π\Omega/2\pi of tens of MHz, ωZX/2π\omega_{ZX}/2\pi is of order 1–2 MHz, giving gate times of hundreds of ns.

Question 3

Which terms does the echoed sequence CR(+) X1 CR(−) X1\mathrm{CR}(+)\,X_1\,\mathrm{CR}(-)\,X_1 cancel to first order?

  1. ZXZX only
  2. IXIX, ZIZI and the static ZZZZ
  3. All terms, leaving the identity
  4. Only terms acting on the target
Show the answer

B. IXIX, ZIZI and the static ZZZZ

Terms odd in Ω\Omega flip with the drive sign; terms containing Z1Z_1 flip under the control's π\pi pulses. ZXZX flips twice and survives; IXIX (odd, no Z1Z_1), and ZIZI and ZZZZ (even, with Z1Z_1), flip once and cancel.

Question 4

How does a tunable coupler placed above both qubit frequencies switch the qubit–qubit interaction off?

  1. By absorbing the photons the qubits emit
  2. By detuning the qubits from each other
  3. By shorting the junctions
  4. The virtual path through the coupler has the opposite sign to the direct capacitive coupling, and at one coupler frequency the two cancel
Show the answer

D. The virtual path through the coupler has the opposite sign to the direct capacitive coupling, and at one coupler frequency the two cancel

g~≈g12+g1g22(1Δ1+1Δ2)\tilde g\approx g_{12}+\frac{g_1g_2}{2}\big(\frac{1}{\Delta_1}+\frac{1}{\Delta_2}\big) with Δj<0\Delta_j<0: the second term is negative and can cancel g12>0g_{12}>0. Tuning ωc\omega_c moves smoothly between off and strongly on.

Further reading

  • C. Rigetti & M. Devoret, "Fully microwave-tunable universal gates in superconducting qubits with linear couplings and fixed transition frequencies", Physical Review B 81, 134507 (2010); J. M. Chow et al., "Simple all-microwave entangling gate for fixed-frequency superconducting qubits", Physical Review Letters 107, 080502 (2011).
  • E. Magesan & J. M. Gambetta, "Effective Hamiltonian models of the cross-resonance gate", Physical Review A 101, 052308 (2020).
  • F. W. Strauch et al., "Quantum logic gates for coupled superconducting phase qubits", Physical Review Letters 91, 167005 (2003); L. DiCarlo et al., "Demonstration of two-qubit algorithms with a superconducting quantum processor", Nature 460, 240 (2009).
  • F. Yan et al., "Tunable coupling scheme for implementing high-fidelity two-qubit gates", Physical Review Applied 10, 054062 (2018).
  • P. Krantz et al., "A quantum engineer's guide to superconducting qubits", Applied Physics Reviews 6, 021318 (2019).