Simple Algorithm
Quantum Computing · Rydberg State · 28 min

Dynamical Decoupling & Noise Spectroscopy

Echoes, filter functions and fighting slow noise.

Much of the dephasing in solid-state qubits comes from noise that is slow, and slow noise can be undone: flip the qubit halfway through and the phase gained in the first half unwinds in the second. This lesson turns that trick into a quantitative theory, with filter functions that predict decoherence from the noise spectrum, pulse sequences that suppress it, and spectroscopy that makes the qubit a noise analyzer.

What you’ll learn

  • Derive the coherence decay χ(T)\chi(T) of a qubit under Gaussian dephasing noise in terms of a filter function.
  • Compute filter functions for free evolution, the Hahn echo and CPMG, and explain their low-frequency behavior.
  • Predict how T2T_2 scales with the number of pulses for power-law noise such as 1/f1/f.
  • Explain dynamical decoupling as group averaging and state its limitations.

Dephasing as a filtering problem

Model pure dephasing by a classical, stationary, zero-mean Gaussian frequency fluctuation β(t)\beta(t), with H(t)=12β(t)ZH(t) = \tfrac12\beta(t)Z. Its statistics are fixed by the correlation function C(τ)=⟨β(t+τ)β(t)⟩C(\tau) = \langle\beta(t+\tau)\beta(t)\rangle, or equivalently by the spectral density S(ω)=∫dτ e−iωτC(τ)S(\omega) = \int d\tau\,e^{-i\omega\tau}C(\tau). Ideal, instantaneous π\pi pulses flip the sign with which the noise acts (since XZX=−ZXZX = -Z); record those signs in a switching function y(t)=±1y(t) = \pm1.

DerivationCoherence from the noise spectrum
  1. φ(T)=∫0Tdt  y(t) β(t)\varphi(T) = \int_0^Tdt\;y(t)\,\beta(t)

    The relative phase accumulated by time TT:

  2. χ(T)=12⟨φ2⟩=12∫0T ⁣ ⁣∫0Tdt dt′  y(t) y(t′) C(t−t′)\chi(T) = \tfrac12\langle\varphi^2\rangle = \tfrac12\int_0^T\!\!\int_0^Tdt\,dt'\;y(t)\,y(t')\,C(t-t')

    For Gaussian noise ⟨e−iφ⟩=e−⟨φ2⟩/2\langle e^{-i\varphi}\rangle = e^{-\langle\varphi^2\rangle/2}, so the coherence decays as e−χ(T)e^{-\chi(T)} with

  3. χ(T)=12π∫0∞dω  S(ω) ∣Y~T(ω)∣2=1π∫0∞dω  S(ω) F(ωT)ω2\chi(T) = \frac{1}{2\pi}\int_0^\infty d\omega\;S(\omega)\,\abs{\tilde Y_T(\omega)}^2 = \frac{1}{\pi}\int_0^\infty d\omega\;S(\omega)\,\frac{F(\omega T)}{\omega^2}

    Fourier transform: with Y~T(ω)=∫0Tdt y(t) eiωt\tilde Y_T(\omega) = \int_0^Tdt\,y(t)\,e^{i\omega t} and an even spectrum SS,

  4. FFID(z)=2sin⁡2z2,Fecho(z)=8sin⁡4z4F_{\text{FID}}(z) = 2\sin^2\frac z2,\qquad F_{\text{echo}}(z) = 8\sin^4\frac z4

    The filter function F(ωT)=12ω2∣Y~T(ω)∣2F(\omega T) = \tfrac12\omega^2\abs{\tilde Y_T(\omega)}^2 says which frequencies the sequence lets through. For free evolution (Ramsey) and for a Hahn echo:

Look at low frequencies. FFID(z)/z2→12F_{\text{FID}}(z)/z^2\to\tfrac12, so quasi-static noise passes undiminished: a static offset of variance σ2\sigma^2 gives the Gaussian decay e−σ2T2/2e^{-\sigma^2T^2/2} behind T2∗T_2^*. But Fecho(z)≈z4/32F_{\text{echo}}(z)\approx z^4/32: the echo is a high-pass filter that blocks noise slower than about 1/T1/T. For white noise, S(ω)=S0S(\omega) = S_0, both give exactly χ=S0T/2\chi = S_0T/2. Echoes only help against correlated (colored) noise.

Check your understanding

Under white dephasing noise (S(ω)=S0S(\omega) = S_0 at all frequencies), how does a single Hahn echo change the coherence at total time TT?

  1. It removes the decay completely
  2. It doubles the coherence time
  3. It has no effect: both give χ=S0T/2\chi = S_0T/2
  4. It makes the decay faster
Show the answer

C. It has no effect: both give χ=S0T/2\chi = S_0T/2

Integrating either filter function against a flat spectrum gives the same χ=S0T/2\chi = S_0T/2. White noise is uncorrelated in time, so nothing gained before the pulse can be cancelled after it.

From echo to CPMG and noise spectroscopy

The Carr–Purcell–Meiboom–Gill (CPMG) sequence uses NN equally spaced π\pi pulses at tj=T(j−12)/Nt_j = T(j-\tfrac12)/N. The switching function becomes a square wave of period 2T/N2T/N, so F(ωT)/ω2F(\omega T)/\omega^2 is a band-pass filter centered at

ωN=πNT,χ(T)  ≈  4Tπ2 S(ωN)(N≫1)\omega_N = \frac{\pi N}{T},\qquad \chi(T)\;\approx\;\frac{4T}{\pi^2}\,S(\omega_N)\qquad(N\gg1)
The leading term. The square wave's odd harmonics add smaller contributions from 3ωN,5ωN,…3\omega_N, 5\omega_N,\dots with weights 19,125,…\tfrac19,\tfrac1{25},\dots

Varying NN and TT sweeps the pass band, and each measured χ\chi reads out SS at one frequency: dynamical-decoupling noise spectroscopy (Álvarez & Suter 2011; Bylander et al. 2011). It has mapped the 1/f1/f-like flux and charge noise of superconducting qubits and the nuclear-spin baths of spin qubits.

DerivationHow T₂ scales with the number of pulses
  1. χ(T)≈4Tπ2 A(TπN)α  ∝  T1+αNα\chi(T)\approx\frac{4T}{\pi^2}\,A\Big(\frac{T}{\pi N}\Big)^{\alpha}\;\propto\;\frac{T^{1+\alpha}}{N^{\alpha}}

    Take a power-law spectrum S(ω)=A/ωαS(\omega) = A/\omega^\alpha (α=1\alpha = 1 for 1/f1/f noise) and use the band-pass approximation:

  2. T2(N)  ∝  Nα/(1+α)T_2(N)\;\propto\;N^{\alpha/(1+\alpha)}

    Define T2(N)T_2(N) by χ=1\chi = 1:

  3. χFID≈AT22πln⁡1ωirT,χecho=ln⁡22π AT2\chi_{\text{FID}}\approx\frac{AT^2}{2\pi}\ln\frac{1}{\omega_{\text{ir}}T},\qquad \chi_{\text{echo}} = \frac{\ln2}{2\pi}\,AT^2

    For 1/f1/f noise, T2∝NT_2\propto\sqrt N and the decay is Gaussian-like (χ∝T2\chi\propto T^2). For a single echo the exact filter gives a closed form, free of the infrared divergence that afflicts the Ramsey decay (cut off at ωir\omega_{\text{ir}}):

Predict before you look

A qubit is limited by 1/f1/f dephasing noise. In the regime where CPMG acts as a narrow band-pass filter, you increase the number of pulses from N=4N = 4 to N=64N = 64. By what factor does T2T_2 grow?

Show the answer

4

T2∝Nα/(1+α)=N1/2T_2\propto N^{\alpha/(1+\alpha)} = N^{1/2} for α=1\alpha = 1, so a 16-fold increase in NN gives 16=4\sqrt{16} = 4. For white noise (α=0\alpha = 0) the factor would be 1; for nearly static noise (α→∞\alpha\to\infty) it would approach 16.

Dynamical decoupling as group averaging

Echoes generalize beyond classical dephasing. Let a qubit couple to a quantum bath through all three Paulis, H=I⊗HB+∑α=x,y,zσα⊗BαH = \id\otimes H_B+\sum_{\alpha = x,y,z}\sigma_\alpha\otimes B_\alpha. Apply pulses so that over one cycle the system spends equal short times τ\tau in each "toggling frame" gg of a group G\mathcal{G} of unitaries acting on the qubit.

DerivationAverage Hamiltonian of a decoupling cycle (Viola, Knill & Lloyd 1999)
  1. U(Tc)=∏g∈Ge−i g†Hg τ=exp⁡ ⁣[−iTc(Hˉ(0)+Hˉ(1)+⋯)],Hˉ(0)=1∣G∣∑g∈Gg†HgU(T_c) = \prod_{g\in\mathcal{G}}e^{-i\,g^\dagger Hg\,\tau} = \exp\!\Big[-iT_c\big(\bar H^{(0)}+\bar H^{(1)}+\cdots\big)\Big],\qquad \bar H^{(0)} = \frac{1}{\abs{\mathcal{G}}}\sum_{g\in\mathcal{G}}g^\dagger Hg

    Each segment evolves by g†e−iHτg=e−ig†Hg τg^\dagger e^{-iH\tau}g = e^{-ig^\dagger Hg\,\tau}. Over a cycle of length Tc=∣G∣τT_c = \abs{\mathcal{G}}\tau, the Magnus expansion gives

  2. 14∑g∈{I,X,Y,Z}g†σαg=0⟹Hˉ(0)=I⊗HB\frac14\sum_{g\in\{\id,X,Y,Z\}}g^\dagger\sigma_\alpha g = 0\qquad\Longrightarrow\qquad\bar H^{(0)} = \id\otimes H_B

    Choose G={I,X,Y,Z}\mathcal{G} = \{\id,X,Y,Z\}. Each Pauli commutes with two group elements and anticommutes with the other two, so the group average of every traceless qubit operator vanishes:

  3. I⏟τ→  X  X⏟τ→  Y  YX∝Z⏟τ→  X  XYX∝Y⏟τ→  Y  YXYX∝I\underbrace{\id}_{\tau}\xrightarrow{\;X\;}\underbrace{X}_{\tau}\xrightarrow{\;Y\;}\underbrace{YX\propto Z}_{\tau}\xrightarrow{\;X\;}\underbrace{XYX\propto Y}_{\tau}\xrightarrow{\;Y\;}YXYX\propto\id

    To first order the qubit is decoupled from the bath; the leading correction Hˉ(1)\bar H^{(1)} is suppressed by a further factor of order TcT_c times the coupling strength. The sequence τ\tau, XX, τ\tau, YY, τ\tau, XX, τ\tau, YY (XY4) visits exactly these frames, up to phases:

Refinements cancel higher orders. Concatenated DD nests sequences recursively (Khodjasteh & Lidar 2005), and Uhrig DD places NN pulses at tj=Tsin⁡2 ⁣(πj2N+2)t_j = T\sin^2\!\big(\tfrac{\pi j}{2N+2}\big), which cancels the pure-dephasing phase up to order NN in TT (Uhrig 2007).

Key takeaways

  • Under Gaussian dephasing noise, coherence decays as e−χ(T)e^{-\chi(T)} with χ=1π∫0∞dω S(ω)F(ωT)/ω2\chi = \tfrac1\pi\int_0^\infty d\omega\,S(\omega)F(\omega T)/\omega^2: the pulse sequence sets the filter FF.
  • Free evolution passes low frequencies (T2∗T_2^*), an echo is a high-pass filter, and CPMG with NN pulses is a band-pass filter at πN/T\pi N/T, which enables noise spectroscopy.
  • For S∝1/ωαS\propto1/\omega^\alpha, T2∝Nα/(1+α)T_2\propto N^{\alpha/(1+\alpha)}; for 1/f1/f noise T2∝NT_2\propto\sqrt N, and white noise is not helped at all.
  • DD is group averaging: XY4 cancels arbitrary single-qubit system–bath coupling to first order, but no sequence beats the relaxation limit T2≤2T1T_2\le2T_1.

Check yourself

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

Question 1

Which noise does a single Hahn echo cancel exactly?

  1. White noise
  2. A frequency offset that is constant during each run but random from run to run
  3. Energy relaxation
  4. Noise at the qubit's transition frequency
Show the answer

B. A frequency offset that is constant during each run but random from run to run

A static offset δ\delta adds phase δT/2\delta T/2 before the π\pi pulse and effectively −δT/2-\delta T/2 after it, for any δ\delta. That is why an echo converts the fast T2∗T_2^* decay caused by shot-to-shot offsets into the slower echo decay.

Question 2

A CPMG sequence with N=20N = 20 pulses over a total time T=40 μsT = 40\,\mu\text{s} is most sensitive to dephasing noise near which angular frequency?

  1. ω≈0\omega\approx0
  2. ω≈π/2 rad/μs\omega\approx\pi/2\ \text{rad}/\mu\text{s}
  3. ω≈40π rad/μs\omega\approx40\pi\ \text{rad}/\mu\text{s}
  4. ω≈π/40 rad/μs\omega\approx\pi/40\ \text{rad}/\mu\text{s}
Show the answer

B. ω≈π/2 rad/μs\omega\approx\pi/2\ \text{rad}/\mu\text{s}

The pulses make y(t)y(t) a square wave with half-period T/N=2 μsT/N = 2\,\mu\text{s}, so the pass band sits at ωN=πN/T=π/2\omega_N = \pi N/T = \pi/2 rad/μs, a frequency of 0.250.25 MHz.

Question 3

A qubit couples to its bath through XX, YY and ZZ simultaneously. Why does the XY4 cycle decouple it to first order, while a train of XX pulses alone does not?

  1. Because XY4 pulses are shorter
  2. Because XY4 averages over the full group {I,X,Y,Z}\{\id,X,Y,Z\}, which kills every traceless qubit operator, while XX pulses average only over {I,X}\{\id,X\} and leave the σx⊗Bx\sigma_x\otimes B_x coupling intact
  3. Because a train of XX pulses does not suppress any noise
  4. Because XY4 deposits more energy into the bath
Show the answer

B. Because XY4 averages over the full group {I,X,Y,Z}\{\id,X,Y,Z\}, which kills every traceless qubit operator, while XX pulses average only over {I,X}\{\id,X\} and leave the σx⊗Bx\sigma_x\otimes B_x coupling intact

With G={I,X}\mathcal{G} = \{\id,X\}, 12(σx+XσxX)=σx\tfrac12(\sigma_x+X\sigma_xX) = \sigma_x survives while σy\sigma_y and σz\sigma_z average to zero. The four-element Pauli group removes all three couplings.

Question 4

Which limit can dynamical decoupling not overcome, however fast the pulses?

  1. The Ramsey time T2∗T_2^*
  2. Dephasing from 1/f1/f noise
  3. The bound T2≤2T1T_2\le2T_1 set by energy relaxation at the qubit frequency
  4. Quasi-static inhomogeneous broadening
Show the answer

C. The bound T2≤2T1T_2\le2T_1 set by energy relaxation at the qubit frequency

Relaxation is driven by noise at the qubit frequency (GHz for superconducting qubits), far above any pulse rate. DD can remove low-frequency dephasing and bring T2T_2 toward 2T12T_1, but not beyond.

Further reading

  • E. L. Hahn, "Spin echoes", Physical Review 80, 580 (1950).
  • H. Y. Carr & E. M. Purcell, Physical Review 94, 630 (1954); S. Meiboom & D. Gill, Review of Scientific Instruments 29, 688 (1958): the CPMG sequence.
  • L. Viola, E. Knill & S. Lloyd, "Dynamical decoupling of open quantum systems", Physical Review Letters 82, 2417 (1999).
  • Ł. Cywiński, R. M. Lutchyn, C. P. Nave & S. Das Sarma, "How to enhance dephasing time in superconducting qubits", Physical Review B 77, 174509 (2008).
  • J. Bylander et al., "Noise spectroscopy through dynamical decoupling with a superconducting flux qubit", Nature Physics 7, 565 (2011).