Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Rabi Oscillations & the Rotating Frame

How a drive field rotates a qubit, and the rotating-wave approximation.

Shine a resonant field on a two-level system and its population sloshes back and forth at a rate set by the field's strength, not its frequency. Making that precise takes a change of frame and an approximation. This lesson examines both: when the rotating-wave approximation holds, what it misses, and how it becomes the pulse toolkit of quantum computers.

What you’ll learn

  • Derive the rotating-frame Hamiltonian H~=ℏ2(ΔZ+Ωcos⁡ϕ X+Ωsin⁡ϕ Y)\tilde H=\tfrac\hbar2(\Delta Z+\Omega\cos\phi\,X+\Omega\sin\phi\,Y) and explain how the drive phase selects the rotation axis.
  • Solve the Rabi problem within the RWA and derive the quadratic error scaling of π pulses under detuning and amplitude errors.
  • Derive the Bloch–Siegert shift Ω2/(4ω0)\Omega^2/(4\omega_0) as the leading correction to the RWA.
  • Analyze Ramsey interferometry and its use for frequency calibration.

The rotating frame

DerivationDriven qubit in the frame of the drive
  1. H(t)=−ℏω02Z+ℏΩcos⁡(ωt−ϕ) XH(t)=-\frac{\hbar\omega_0}{2}Z+\hbar\Omega\cos(\omega t-\phi)\,X

    Qubit with ground state ∣0⟩\ket0, driven at frequency ω\omega with phase ϕ\phi and strength Ω\Omega.

  2. V=e−iωtZ/2:VXV†=cos⁡(ωt)X+sin⁡(ωt)Y,iℏV˙V†=ℏω2ZV=e^{-i\omega tZ/2}:\quad VXV^\dagger=\cos(\omega t)X+\sin(\omega t)Y,\qquad i\hbar\dot VV^\dagger=\tfrac{\hbar\omega}{2}Z

    Rotate with the drive.

  3. H~=ℏΔ2Z+ℏΩ2(cos⁡ϕ X+sin⁡ϕ Y)+ℏΩ2(cos⁡(2ωt−ϕ) X+sin⁡(2ωt−ϕ) Y)\tilde H=\frac{\hbar\Delta}{2}Z+\frac{\hbar\Omega}{2}\big(\cos\phi\,X+\sin\phi\,Y\big)+\frac{\hbar\Omega}{2}\big(\cos(2\omega t-\phi)\,X+\sin(2\omega t-\phi)\,Y\big)

    Product-to-sum formulas split the drive into a static co-rotating part and a counter-rotating part at 2ω2\omega. Here Δ=ω−ω0\Delta=\omega-\omega_0.

  4. H~RWA=ℏ2(Δ Z+Ωcos⁡ϕ X+Ωsin⁡ϕ Y)\tilde H_{\rm RWA}=\frac\hbar2\big(\Delta\,Z+\Omega\cos\phi\,X+\Omega\sin\phi\,Y\big)

    Rotating-wave approximation: drop the counter-rotating part. Amplitude sets the rotation rate and phase sets the axis in the equatorial plane. That is how RxR_x and RyR_y gates are made, and why a ZZ rotation can be done virtually, by shifting the phase of all later pulses (McKay et al., 2017).

The Rabi solution and its error budget

Averaged over time, P1‾=12 Ω2/(Ω2+Δ2)\overline{P_1}=\tfrac12\,\Omega^2/(\Omega^2+\Delta^2): a Lorentzian in Δ\Delta with full width 2Ω2\Omega. Stronger driving broadens the line (power broadening). For gates the more useful question is how sensitive a calibrated pulse is to small errors.

DerivationErrors of a π pulse are quadratic
  1. P1=11+δ2 sin⁡2 ⁣(π21+δ2),δ=ΔΩ,t=πΩP_1=\frac{1}{1+\delta^2}\,\sin^2\!\Big(\frac\pi2\sqrt{1+\delta^2}\Big),\qquad\delta=\frac\Delta\Omega,\quad t=\frac\pi\Omega

    A π pulse calibrated on resonance, applied with a small detuning.

  2. sin⁡2 ⁣(π21+δ2)=cos⁡2 ⁣(π2(1+δ2−1))=1−O(δ4)\sin^2\!\Big(\frac\pi2\sqrt{1+\delta^2}\Big)=\cos^2\!\Big(\frac\pi2\big(\sqrt{1+\delta^2}-1\big)\Big)=1-\order{\delta^4}

    The over-rotation caused by W>ΩW>\Omega is only fourth order.

  3. 1−P1=δ2+O(δ4)1-P_1=\delta^2+\order{\delta^4}

    The tilt of the axis dominates. Δ=0.1 Ω\Delta=0.1\,\Omega leaves about 1%1\% in ∣0⟩\ket0.

  4. Ω→(1+ε)Ω:1−P1=sin⁡2 ⁣(πε2)≈(πε2)2\Omega\to(1+\varepsilon)\Omega:\qquad1-P_1=\sin^2\!\big(\tfrac{\pi\varepsilon}2\big)\approx\big(\tfrac{\pi\varepsilon}{2}\big)^2

    An amplitude (or duration) error ε\varepsilon also enters quadratically. Composite pulses cancel these terms to higher order.

Beyond the RWA: the Bloch–Siegert shift

The counter-rotating term has strength Ω/2\Omega/2 and oscillates at 2ω2\omega. To first order it averages to zero; its leading effect is second order, and it is a frequency shift.

DerivationLeading high-frequency correction
  1. H~(t)=H~RWA+h e−2iωt+h†e2iωt,h=ℏΩ2 σ+,σ+=∣0⟩ ⁣⟨1∣\tilde H(t)=\tilde H_{\rm RWA}+h\,e^{-2i\omega t}+h^\dagger e^{2i\omega t},\qquad h=\tfrac{\hbar\Omega}{2}\,\sigma_+,\quad\sigma_+=\ketbra{0}{1}

    Write the counter-rotating part (ϕ=0\phi=0) with σ+=12(X+iY)\sigma_+=\tfrac12(X+iY).

  2. H~eff=H~RWA+[h†,h]2ℏω+…\tilde H_{\rm eff}=\tilde H_{\rm RWA}+\frac{[h^\dagger,h]}{2\hbar\omega}+\dots

    Leading order of the high-frequency (Floquet–Magnus) expansion: for H0+h e−iνt+h†eiνtH_0+h\,e^{-i\nu t}+h^\dagger e^{i\nu t} with ν\nu large compared with ∥H0∥/ℏ\norm{H_0}/\hbar and ∥h∥/ℏ\norm h/\hbar, the period-averaged dynamics is generated by H0+[h†,h]/(ℏν)H_0+[h^\dagger,h]/(\hbar\nu), with corrections of higher order in 1/ν1/\nu.

  3. [σ−,σ+]=−Z  ⟹  H~eff=ℏ2(Δ−Ω24ω)Z+ℏΩ2X[\sigma_-,\sigma_+]=-Z\;\Longrightarrow\;\tilde H_{\rm eff}=\frac\hbar2\Big(\Delta-\frac{\Omega^2}{4\omega}\Big)Z+\frac{\hbar\Omega}{2}X

    Here [h†,h]=ℏ2Ω24[σ−,σ+]=−ℏ2Ω24Z[h^\dagger,h]=\tfrac{\hbar^2\Omega^2}{4}[\sigma_-,\sigma_+]=-\tfrac{\hbar^2\Omega^2}{4}Z.

  4. ωres≈ω0+Ω24ω0(Bloch–Siegert shift)\omega_{\rm res}\approx\omega_0+\frac{\Omega^2}{4\omega_0}\qquad\text{(Bloch--Siegert shift)}

    Resonance now needs Δ=Ω2/(4ω)\Delta=\Omega^2/(4\omega), so the line moves up by δBS≈Ω2/(4ω0)\delta_{\rm BS}\approx\Omega^2/(4\omega_0) (Bloch & Siegert, 1940). Physically: virtual counter-rotating processes, detuned by about 2ω02\omega_0, push the two levels apart.

Check your understanding

A qubit at ω0/2π=5\omega_0/2\pi=5 GHz is driven with Ω/2π=50\Omega/2\pi=50 MHz. Roughly how large is the Bloch–Siegert shift δBS/2π\delta_{\rm BS}/2\pi?

  1. 500 kHz
  2. 12.5 MHz
  3. 125 kHz
  4. 785 kHz
Show the answer

C. 125 kHz

In frequency units (50 MHz)2/(4×5000 MHz)=0.125(50\text{ MHz})^2/(4\times5000\text{ MHz})=0.125 MHz =125=125 kHz: tiny compared with Ω\Omega, but larger than the linewidth of a good qubit.

Ramsey interferometry

Two π/2\pi/2 pulses separated by free evolution make an interferometer on the Bloch sphere. Between pulses the drive is off, and in the rotating frame a detuned qubit precesses about zz at rate Δ\Delta.

DerivationRamsey fringes
  1. ∣0⟩→  Rx(π/2)  12(∣0⟩−i∣1⟩),r⃗=−y^\ket0\xrightarrow{\;R_x(\pi/2)\;}\tfrac1{\sqrt2}(\ket0-i\ket1),\qquad\vec r=-\hat y

    First pulse, assumed short compared with 1/∣Δ∣1/\abs\Delta.

  2. r⃗(τ)=(sin⁡Δτ, −cos⁡Δτ, 0)\vec r(\tau)=\big(\sin\Delta\tau,\ -\cos\Delta\tau,\ 0\big)

    Free evolution under H~=ℏ2ΔZ\tilde H=\tfrac\hbar2\Delta Z: precession by Δτ\Delta\tau about zz.

  3. Rx(π2):(x,y,z)↦(x,−z,y)  ⟹  r⃗=(sin⁡Δτ, 0, −cos⁡Δτ)R_x(\tfrac\pi2):(x,y,z)\mapsto(x,-z,y)\;\Longrightarrow\;\vec r=\big(\sin\Delta\tau,\ 0,\ -\cos\Delta\tau\big)

    Second pulse.

  4. P1(τ)=12(1−rz)=cos⁡2Δτ2P_1(\tau)=\tfrac12(1-r_z)=\cos^2\frac{\Delta\tau}{2}

    The fringes oscillate at the detuning; at Δ=0\Delta=0 the two pulses add to a π pulse. With dephasing, the fringe contrast decays on the timescale T2∗T_2^*. Fitting the fringe frequency measures Δ\Delta, which is how qubit frequencies are calibrated.

Predict before you look

In a Ramsey experiment the drive is detuned by Δ/2π=1\Delta/2\pi=1 MHz. At what free-evolution time τ\tau (in ns) does P1P_1 first reach 0?

Show the answer

500 ns

P1=cos⁡2(Δτ/2)=0P_1=\cos^2(\Delta\tau/2)=0 first at Δτ=π\Delta\tau=\pi, so τ=π/(2π×106 s−1)=0.5 μ\tau=\pi/(2\pi\times10^6\,\mathrm{s^{-1}})=0.5\ \mus. The fringe period is 1/(1 MHz)=1 μ1/(1\text{ MHz})=1\ \mus. With NN repetitions of length τ\tau, the frequency is resolved to about 1/(τN)1/(\tau\sqrt N): longer τ\tau helps, up to the coherence time.

Key takeaways

  • In the frame rotating at the drive frequency, the fast ω0\omega_0 precession disappears and the drive splits into a static co-rotating part and a counter-rotating part at 2ω2\omega.
  • Within the RWA, H~=ℏ2(ΔZ+Ωcos⁡ϕ X+Ωsin⁡ϕ Y)\tilde H=\tfrac\hbar2(\Delta Z+\Omega\cos\phi\,X+\Omega\sin\phi\,Y): amplitude sets the rate, phase sets the axis, detuning tilts it, and P1=Ω2W2sin⁡2Wt2P_1=\tfrac{\Omega^2}{W^2}\sin^2\tfrac{Wt}2.
  • Pulse errors are quadratic: 1−P1≈(Δ/Ω)21-P_1\approx(\Delta/\Omega)^2 for detuning and ≈(πε/2)2\approx(\pi\varepsilon/2)^2 for amplitude error.
  • The leading non-RWA effect is the Bloch–Siegert shift Ω2/(4ω0)\Omega^2/(4\omega_0). Ramsey fringes cos⁡2(Δτ/2)\cos^2(\Delta\tau/2) measure detunings and T2∗T_2^*.

Check yourself

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

Question 1

With the same resonant drive, how do you implement Ry(π/2)R_y(\pi/2) instead of Rx(π/2)R_x(\pi/2)?

  1. Double the pulse duration
  2. Shift the drive phase by 90∘90^\circ
  3. Detune the drive by Ω\Omega
  4. Use only the counter-rotating component
Show the answer

B. Shift the drive phase by 90∘90^\circ

On resonance H~RWA=ℏΩ2(cos⁡ϕ X+sin⁡ϕ Y)\tilde H_{\rm RWA}=\tfrac{\hbar\Omega}{2}(\cos\phi\,X+\sin\phi\,Y), so ϕ=π/2\phi=\pi/2 gives the YY axis. The same envelope with a shifted phase rotates about a different equatorial axis.

Question 2

Why are the populations P0,P1P_0,P_1 the same in the lab frame and in the frame defined by V=e−iωtZ/2V=e^{-i\omega tZ/2}?

  1. Because the RWA is exact for populations
  2. Because the drive is weak
  3. Because VV is diagonal in the computational basis, so it changes only the phases of the amplitudes
  4. They are not; populations must be transformed back
Show the answer

C. Because VV is diagonal in the computational basis, so it changes only the phases of the amplitudes

VV multiplies the ∣0⟩\ket0 and ∣1⟩\ket1 amplitudes by e∓iωt/2e^{\mp i\omega t/2} and leaves their moduli unchanged. Only the azimuth of the Bloch vector differs between frames.

Question 3

A π pulse calibrated on resonance is applied with detuning Δ=0.1 Ω\Delta=0.1\,\Omega. About how much population is left in ∣0⟩\ket0?

  1. 10−210^{-2}
  2. 10−110^{-1}
  3. 10−410^{-4}
  4. 00: π pulses are insensitive to detuning
Show the answer

A. 10−210^{-2}

1−P1=δ2+O(δ4)1-P_1=\delta^2+\order{\delta^4} with δ=Δ/Ω\delta=\Delta/\Omega, so about 10−210^{-2} (numerically 0.01000.0100).

Question 4

In a Ramsey sequence (Rx(π/2)R_x(\pi/2), free evolution for τ\tau, Rx(π/2)R_x(\pi/2)) with detuning Δ\Delta and no decoherence, P1(τ)P_1(\tau) equals:

  1. sin⁡2(Ωτ/2)\sin^2(\Omega\tau/2)
  2. e−τ/T2e^{-\tau/T_2}
  3. 12\tfrac12 for all τ\tau
  4. cos⁡2(Δτ/2)\cos^2(\Delta\tau/2)
Show the answer

D. cos⁡2(Δτ/2)\cos^2(\Delta\tau/2)

Between the pulses the Bloch vector precesses about zz by Δτ\Delta\tau, and the second pulse maps that azimuth onto rz=−cos⁡Δτr_z=-\cos\Delta\tau, so P1=cos⁡2(Δτ/2)P_1=\cos^2(\Delta\tau/2).

Further reading

  • I. I. Rabi, Space quantization in a gyrating magnetic field, Physical Review (1937).
  • F. Bloch & A. Siegert, Magnetic resonance for nonrotating fields, Physical Review (1940).
  • N. F. Ramsey, A molecular beam resonance method with separated oscillating fields, Physical Review (1950).
  • D. C. McKay, C. J. Wood, S. Sheldon, J. M. Chow & J. M. Gambetta, Efficient Z gates for quantum computing, Physical Review A (2017).
  • P. Krantz et al., A quantum engineer's guide to superconducting qubits, Applied Physics Reviews (2019).