Simple Algorithm
Quantum Computing · Rydberg State · 26 min

Leakage & Non-Ideal Two-Level Systems

Anharmonicity, leakage out of the qubit subspace, and DRAG pulses.

A transmon is not a two-level system. It is a ladder whose rung spacing shrinks by only a few percent per step, so a fast pulse aimed at 0↔10\leftrightarrow1 also tickles 1↔21\leftrightarrow2. This lesson quantifies that leakage, explains why it is worse than an ordinary error, and derives DRAG, the pulse-shaping trick that cancels it to leading order.

What you’ll learn

  • Model a transmon as a weakly anharmonic (Kerr) ladder and state the asymptotic trade-off between anharmonicity and charge dispersion.
  • Estimate leakage from the spectrum of the pulse envelope at the 1→21\to2 detuning, and explain the accompanying phase errors.
  • Derive the first-order DRAG condition Ωy=−Ω˙x/α\Omega_y=-\dot\Omega_x/\alpha and show that it leaves the qubit rotation unchanged.
  • Explain why leakage is especially damaging for error correction, and how it is measured (L1L_1, L2L_2) and removed.

A ladder, not a pair

A transmon is an LC oscillator whose inductor is a Josephson junction. Expanding the junction's cosine potential beyond quadratic order gives, to leading order, a Kerr (Duffing) oscillator:

H=ℏω01 b†b+ℏα2 b†b†b b,En=ℏω01 n+ℏα2 n(n−1)H=\hbar\omega_{01}\,b^\dagger b+\frac{\hbar\alpha}{2}\,b^\dagger b^\dagger b\,b,\qquad E_n=\hbar\omega_{01}\,n+\frac{\hbar\alpha}{2}\,n(n-1)
Transitions: ω01\omega_{01}, then ω12=ω01+α\omega_{12}=\omega_{01}+\alpha, ω23=ω01+2α\omega_{23}=\omega_{01}+2\alpha, and so on. For a transmon α<0\alpha<0, typically α/2π≈−(200\alpha/2\pi\approx-(200–300)300) MHz against ω01/2π≈4\omega_{01}/2\pi\approx4–66 GHz. Charge matrix elements grow like a harmonic oscillator's, ∣⟨n+1|b†|n⟩∣=n+1\abs{\mel{n+1}{b^\dagger}{n}}=\sqrt{n+1}, so the 1→21\to2 coupling is λ≈2\lambda\approx\sqrt2 times the 0→10\to1 coupling.
Predict before you look

A transmon has EJ/EC=50E_J/E_C=50. Using the asymptotic formula, what is the magnitude of its relative anharmonicity ∣α∣/ω01\abs{\alpha}/\omega_{01}, in percent?

Show the answer

5 %

(8⋅50)−1/2=1/400=1/20=5%(8\cdot50)^{-1/2}=1/\sqrt{400}=1/20=5\%. At ω01/2π=5\omega_{01}/2\pi=5 GHz that is α/2π≈−250\alpha/2\pi\approx-250 MHz: the 1→21\to2 line sits only 250250 MHz below the qubit line.

Leakage: the spectral picture

Drive at ω01\omega_{01} and move to the rotating frame, keeping three levels. With complex envelope ε(t)=Ωx(t)+iΩy(t)\varepsilon(t)=\Omega_x(t)+i\Omega_y(t), the RWA Hamiltonian is

H~=ℏα∣2⟩ ⁣⟨2∣+ℏ2[ε(t)(∣1⟩ ⁣⟨0∣+λ∣2⟩ ⁣⟨1∣)+ε∗(t)(∣0⟩ ⁣⟨1∣+λ∣1⟩ ⁣⟨2∣)]\tilde H=\hbar\alpha\dyad{2}+\frac\hbar2\Big[\varepsilon(t)\big(\ketbra{1}{0}+\lambda\ketbra{2}{1}\big)+\varepsilon^*(t)\big(\ketbra{0}{1}+\lambda\ketbra{1}{2}\big)\Big]
On {∣0⟩,∣1⟩}\{\ket0,\ket1\} this is ℏ2(ΩxX+ΩyY)\tfrac\hbar2(\Omega_xX+\Omega_yY), the Rabi lesson's drive. The extra terms couple ∣1⟩\ket1 to ∣2⟩\ket2, which is detuned by α\alpha from the drive.
DerivationFirst-order leakage estimate
  1. c2(T)≈−iλ2∫0Tε(t) c1(t) eiαt dtc_2(T)\approx-\frac{i\lambda}{2}\int_0^T\varepsilon(t)\,c_1(t)\,e^{i\alpha t}\,dt

    First-order perturbation theory in the ∣2⟩ ⁣⟨1∣\ketbra{2}{1} coupling, in the interaction picture with respect to ℏα∣2⟩ ⁣⟨2∣\hbar\alpha\dyad{2}.

  2. Pleak∼λ24 ∣ε^(α)∣2,ε^(ν)=∫ε(t) eiνt dtP_{\rm leak}\sim\frac{\lambda^2}{4}\,\abs{\hat\varepsilon(\alpha)}^2,\qquad\hat\varepsilon(\nu)=\int\varepsilon(t)\,e^{i\nu t}\,dt

    For an estimate, treat c1(t)c_1(t) as an O(1)\order1 constant: leakage is set by the envelope's spectral weight at the leakage detuning.

  3. ε(t)=θσ2π e−t2/2σ2  ⟹  Pleak∼λ2θ24 e−α2σ2\varepsilon(t)=\frac{\theta}{\sigma\sqrt{2\pi}}\,e^{-t^2/2\sigma^2}\;\Longrightarrow\;P_{\rm leak}\sim\frac{\lambda^2\theta^2}{4}\,e^{-\alpha^2\sigma^2}

    A Gaussian of rotation angle (area) θ\theta has ε^(α)=θ e−α2σ2/2\hat\varepsilon(\alpha)=\theta\,e^{-\alpha^2\sigma^2/2}.

  4. With ∣α∣/2π=200\abs\alpha/2\pi=200 MHz and σ=1\sigma=1 ns, (ασ)2≈1.6(\alpha\sigma)^2\approx1.6 and the suppression factor is only ≈0.2\approx0.2: a short π pulse leaks badly. At σ=3\sigma=3 ns the factor is below 10−610^{-6}. Real pulses are truncated, which adds spectral tails, so practical gates use tg≳4σt_g\gtrsim4\sigma plus the corrections below.

Check your understanding

Why is leakage more harmful to error correction than a bit flip of the same probability?

  1. It only happens when the qubit is idle
  2. A leaked qubit is outside the code space: syndrome circuits misbehave, it can persist for many rounds, and it spreads correlated errors to its neighbors
  3. It always causes an immediate logical error
  4. It cannot be detected, even in principle
Show the answer

B. A leaked qubit is outside the code space: syndrome circuits misbehave, it can persist for many rounds, and it spreads correlated errors to its neighbors

Codes are built for errors that act within the qubit space, where syndrome measurements digitize them into Pauli errors. A transmon in ∣2⟩\ket2 makes the stabilizer circuits around it behave erratically. It lives for about T1T_1, and through two-qubit gates it generates correlated errors, which codes handle poorly. Hence leakage-reduction units and multi-level reset.

DRAG: cancelling the leakage line

DerivationSpectral proof sketch
  1. f˙^(ν)=∫f˙(t) eiνt dt=−iν f^(ν)\widehat{\dot f}(\nu)=\int\dot f(t)\,e^{i\nu t}\,dt=-i\nu\,\hat f(\nu)

    Integrate by parts; the boundary terms vanish because the envelope does.

  2. ε^(ν)=Ω^x(ν)+i(−1α)(−iν) Ω^x(ν)=(1−να)Ω^x(ν)\hat\varepsilon(\nu)=\hat\Omega_x(\nu)+i\Big(-\frac1\alpha\Big)(-i\nu)\,\hat\Omega_x(\nu)=\Big(1-\frac{\nu}{\alpha}\Big)\hat\Omega_x(\nu)

    Transform the complex envelope ε=Ωx+iΩy\varepsilon=\Omega_x+i\Omega_y with the DRAG quadrature.

  3. ε^(α)=0,ε^(0)=Ω^x(0)=θ,∫Ωy dt=−Ωx(T)−Ωx(0)α=0\hat\varepsilon(\alpha)=0,\qquad\hat\varepsilon(0)=\hat\Omega_x(0)=\theta,\qquad\int\Omega_y\,dt=-\frac{\Omega_x(T)-\Omega_x(0)}{\alpha}=0

    The spectrum has an exact zero at the leakage line. The qubit line (ν=0\nu=0) still sees area θ\theta, and the quadrature has no net area, so there is no extra YY rotation.

  4. The spectral argument freezes c1(t)c_1(t). The systematic derivation uses a time-dependent frame that dresses the qubit levels with ∣2⟩\ket2 order by order in 1/α1/\alpha; it confirms the quadrature and supplies the detuning correction. First-order variants that target leakage or phase errors differ in the quadrature coefficient by about a factor of two, so experiments write Ωy=−β Ω˙x/α\Omega_y=-\beta\,\dot\Omega_x/\alpha and calibrate β\beta. □\square

python
import numpy as np

def drag_pulse(t, t0, sigma, theta, alpha, beta=1.0):
    """Gaussian in-phase envelope of area theta, plus the DRAG quadrature."""
    A = theta / (sigma * np.sqrt(2 * np.pi))
    omega_x = A * np.exp(-((t - t0) ** 2) / (2 * sigma**2))
    domega_x = -(t - t0) / sigma**2 * omega_x
    omega_y = -beta * domega_x / alpha
    return omega_x, omega_y  # rad/ns when t, sigma are in ns and alpha in rad/ns

t = np.linspace(0.0, 20.0, 401)  # a 20 ns gate, sigma = 5 ns
ox, oy = drag_pulse(t, t0=10.0, sigma=5.0, theta=np.pi, alpha=-2 * np.pi * 0.25)

A DRAG π pulse for a transmon with α/2π=−250\alpha/2\pi=-250 MHz. The quadrature is an antisymmetric "derivative" lobe. Truncating at ±2σ\pm2\sigma leaves edge jumps of about 14%14\% of the peak (e−2e^{-2}), so real waveforms subtract that offset, making the envelope start and end at zero as the derivation assumes.

Measuring and managing leakage

Leakage randomized benchmarking tracks the population outside {∣0⟩,∣1⟩}\{\ket0,\ket1\} after mm random gates and fits it to a two-state Markov model. Per gate, population leaks out with probability L1L_1 and seeps back with probability L2L_2 (Wood & Gambetta, 2018):

pleak(m)=L1L1+L2(1−(1−L1−L2)m)p_{\rm leak}(m)=\frac{L_1}{L_1+L_2}\Big(1-(1-L_1-L_2)^m\Big)
Leaked population saturates at L1/(L1+L2)L_1/(L_1+L_2). Fitting the approach gives L1L_1 and L2L_2 separately from the ordinary error rate.
Non-idealityOriginTypical mitigation
Leakage to ∣2⟩,∣3⟩,…\ket2,\ket3,\dotsweak anharmonicity; fast or strong pulsesDRAG, longer pulses, leakage-reduction units, reset
Charge dispersionoffset-charge dependence of levels, ∝e−8EJ/EC\propto e^{-\sqrt{8E_J/E_C}}large EJ/ECE_J/E_C (the transmon regime)
Two-level-system defectsmicroscopic defects in amorphous oxides couple resonantly to the qubit; T1T_1 fluctuates in time and frequencymaterials and fabrication, frequency tuning away from defects
Static ZZZZ couplingrepulsion between ∣11⟩\ket{11} and ∣02⟩,∣20⟩\ket{02},\ket{20} in coupled transmonstunable couplers, echo sequences
Quasiparticles and radiationbroken Cooper pairs; high-energy impacts cause correlated error burstsshielding, phonon and quasiparticle traps
Most of these appear because the "qubit" is two levels of a larger system that is coupled to a larger world.

Key takeaways

  • A transmon is a weakly anharmonic ladder: ω12=ω01+α\omega_{12}=\omega_{01}+\alpha with ℏα≈−EC\hbar\alpha\approx-E_C, a few percent of ω01\omega_{01}. Large EJ/ECE_J/E_C buys exponential charge-noise immunity at an algebraic cost in anharmonicity.
  • Leakage is driven by the envelope's spectral weight at the 1→21\to2 detuning, roughly e−α2σ2e^{-\alpha^2\sigma^2} for Gaussians, and AC Stark phase errors accompany it. Speed and leakage trade off.
  • First-order DRAG, Ωy=−Ω˙x/α\Omega_y=-\dot\Omega_x/\alpha, puts a spectral zero at ν=α\nu=\alpha while preserving the rotation; a detuning correction handles the phase error, and β\beta is calibrated in practice.
  • Leakage takes qubits outside the code space and causes long-lived, correlated errors. It is measured with leakage RB (L1L_1, L2L_2) and removed with leakage-reduction units and reset.

Check yourself

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

Question 1

A transmon has ω01/2π=5\omega_{01}/2\pi=5 GHz and α/2π=−250\alpha/2\pi=-250 MHz. What is the 1→21\to2 transition frequency?

  1. 5.25 GHz
  2. 4.75 GHz
  3. 4.5 GHz
  4. 9.75 GHz
Show the answer

B. 4.75 GHz

ω12=ω01+α\omega_{12}=\omega_{01}+\alpha, so 4.754.75 GHz. A drive at 5 GHz is detuned from it by only 250 MHz, 5% of the qubit frequency.

Question 2

What does the DRAG quadrature Ωy=−Ω˙x/α\Omega_y=-\dot\Omega_x/\alpha do to the spectrum of the complex envelope ε=Ωx+iΩy\varepsilon=\Omega_x+i\Omega_y?

  1. It narrows the whole spectrum
  2. It shifts the spectrum by α\alpha
  3. It places a zero at the leakage detuning ν=α\nu=\alpha while keeping the pulse area at ν=0\nu=0
  4. It removes the counter-rotating terms
Show the answer

C. It places a zero at the leakage detuning ν=α\nu=\alpha while keeping the pulse area at ν=0\nu=0

ε^(ν)=(1−ν/α)Ω^x(ν)\hat\varepsilon(\nu)=(1-\nu/\alpha)\hat\Omega_x(\nu): a zero exactly at ν=α\nu=\alpha, unchanged weight θ\theta at ν=0\nu=0.

Question 3

In the leakage model with leakage rate L1L_1 and seepage rate L2L_2 per gate, what fraction of the population is leaked after very many gates?

  1. L1L_1
  2. L2/(L1+L2)L_2/(L_1+L_2)
  3. 11
  4. L1/(L1+L2)L_1/(L_1+L_2)
Show the answer

D. L1/(L1+L2)L_1/(L_1+L_2)

The update p↦p(1−L2)+(1−p)L1p\mapsto p(1-L_2)+(1-p)L_1 has fixed point p∗=L1/(L1+L2)p^*=L_1/(L_1+L_2), approached geometrically as (1−L1−L2)m(1-L_1-L_2)^m.

Question 4

Raising a transmon's EJ/ECE_J/E_C from 50 to 100 has roughly what effect?

  1. It reduces the relative anharmonicity by about 2\sqrt2 and suppresses charge dispersion by more than three orders of magnitude
  2. It halves both the anharmonicity and the charge dispersion
  3. It increases the anharmonicity
  4. It leaves the charge dispersion unchanged
Show the answer

A. It reduces the relative anharmonicity by about 2\sqrt2 and suppresses charge dispersion by more than three orders of magnitude

∣α∣/ω01≈(8EJ/EC)−1/2\abs\alpha/\omega_{01}\approx(8E_J/E_C)^{-1/2} falls from 5% to about 3.5%. The exponent 8EJ/EC\sqrt{8E_J/E_C} grows from 20 to about 28.3, so e−8EJ/ECe^{-\sqrt{8E_J/E_C}} drops by e−8.3≈2.5×10−4e^{-8.3}\approx2.5\times10^{-4}, partly offset by a slowly growing prefactor.

Further reading

  • J. Koch, T. M. Yu, J. Gambetta, A. A. Houck, D. I. Schuster, J. Majer, A. Blais, M. H. Devoret, S. M. Girvin & R. J. Schoelkopf, Charge-insensitive qubit design derived from the Cooper pair box, Physical Review A (2007).
  • F. Motzoi, J. M. Gambetta, P. Rebentrost & F. K. Wilhelm, Simple pulses for elimination of leakage in weakly nonlinear qubits, Physical Review Letters (2009).
  • J. M. Gambetta, F. Motzoi, S. T. Merkel & F. K. Wilhelm, Analytic control methods for high-fidelity unitary operations in a weakly nonlinear oscillator, Physical Review A (2011).
  • C. J. Wood & J. M. Gambetta, Quantification and characterization of leakage errors, Physical Review A (2018).
  • P. Aliferis & B. M. Terhal, Fault-tolerant quantum computation for local leakage faults, Quantum Information & Computation (2007).