Decoherence Mechanisms in Qubits

Decoherence is the number every quantum-hardware group is measured by, and the past two decades have been a campaign against it. Superconducting-qubit coherence has climbed from the nanoseconds of the first Cooper-pair box in 1999 to hundreds of microseconds in three-dimensional and tantalum-based transmons, driven by attacking the terms in $S(\omega)$: better dielectrics to starve the two-level-system bath, gap engineering to suppress quasiparticles, Purcell filters to close the readout-loss channel, and flux sweet spots to evade $1/f$ noise. The same filter-function machinery that explains spin echo now runs in reverse as noise spectroscopy — sweeping the spacing of dynamical-decoupling pulses reconstructs $S_\lambda(\omega)$ and diagnoses which bath limits a device. And decoherence sets the target for quantum error correction: the surface code only helps once physical error rates, dominated by decoherence during gates and idling, fall below roughly one percent — the threshold the coherence budget of Worked Example 3 makes concrete. The open frontier is the low-frequency noise no one has fully explained: the microscopic origin of $1/f$ flux and charge noise, and of the individual defects that occasionally swallow a qubit's $T_1$, remains an active and humbling problem.

Picture a qubit superposition $\alpha\lvert 0\rangle + \beta\lvert 1\rangle$ as a tiny clock whose hand points somewhere on the equator of the Bloch sphere. Two physically distinct things can go wrong. Relaxation, the $T_1$ process, is the clock's mainspring running down: the qubit emits its energy quantum into the bath and slides back toward the ground state. Dephasing, the $T_2$ process, is the clock running at a slightly wrong, fluctuating rate: the hand's angle becomes uncertain, and averaging over many noisy repetitions washes the equatorial component to zero even though no energy was lost. The first figure puts both decays side by side with real numbers.

The crucial insight is frequency matching. To absorb or emit a quantum $\hbar\omega_{01}$, the bath must have spectral weight at $\omega_{01}$ — so relaxation is set by the noise power at the qubit frequency, $S(\omega_{01})$. Pure dephasing needs no energy exchange; it is driven by slow wobbles of the transition frequency, so it is set by $S(\omega)$ near DC. This is why sluggish $1/f$ noise — enormous weight at low frequency, almost none at gigahertz — dephases a qubit ferociously while barely relaxing it. The second figure unifies the two timescales: the noise spectral density itself, with both sampling points marked.

A sweet spot is the noise-engineering masterstroke this picture makes obvious. If the qubit frequency depends on a noisy control parameter $\lambda$ — a gate charge, a flux — tune the bias to where $\partial\omega_{01}/\partial\lambda = 0$. There the qubit is, to first order, blind to fluctuations of $\lambda$; only the weaker second-order curvature survives. The transmon does this for charge noise at large $E_J/E_C$, the flux qubit at its symmetry point. Everything this article develops is a way to flatten $S(\omega)$ where the qubit reads it, or move the qubit to read a quieter part of the spectrum.

Start with the phenomenology. Independent of any microscopic model, the density matrix of a single qubit subject to relaxation and dephasing evolves so that populations relax at rate $1/T_1$ and coherences decay at rate $1/T_2$ on top of free precession at $\omega_{01}$. This is the two-timescale skeleton that every experiment fits to.

where: $\rho_{ij}$ — density-matrix elements, diagonal entries are populations and off-diagonal entries are coherences; $T_1$ — longitudinal (energy) relaxation time; $T_2$ — transverse (coherence) decay time; $\omega_{01}$ — qubit transition frequency. Reading: the diagonal always sums to one, so the map is trace-preserving; the excited population decays with $T_1$, the superposition-encoding coherence with $T_2$. The two timescales are bound together because energy relaxation itself destroys coherence. Splitting the coherence decay into the part inherited from relaxation and the part from genuine phase noise gives the master relation of the subject.

where: $T_\phi$ — pure-dephasing time, phase loss with no energy exchange. Reading: relaxation feeds the coherence-decay rate with a factor $\tfrac12$, so a perfectly relaxation-limited qubit ($T_\phi\to\infty$) has $T_2 = 2T_1$; any real dephasing only shortens $T_2$. Now the microscopics. Applying Fermi's golden rule to the coupling, the relaxation rate is the squared transition matrix element times the noise power at the qubit frequency — the only frequency at which the bath can absorb the energy quantum.

where: $\hat{A}$ — qubit operator coupling to the noisy variable $\lambda$; $S_\lambda(\omega)=\int \langle\lambda(t)\lambda(0)\rangle\,e^{i\omega t}\,\mathrm dt$ — power spectral density of the noise; $\omega_{01}$ — qubit frequency. Reading: relaxation is resonant — it cares only about the bath's spectral weight at $\omega_{01}$, nothing about DC. Pure dephasing is the complementary limit. Slow noise modulates the qubit splitting without flipping the state, so the relevant quantity is the frequency's sensitivity to the noisy parameter times the noise power near zero frequency.

where: $(\partial \hat H_{\rm sys}/\partial\lambda)_{01}$ — first-order sensitivity of the qubit splitting to $\lambda$; $S_\lambda(\omega\to 0)$ — near-DC noise power. Reading: at a sweet spot the prefactor vanishes and first-order dephasing switches off, leaving only weak second-order curvature — the whole justification for sweet-spot design. When the noise is $1/f$ rather than white, equation (4) fails because $S_\lambda(\omega\to 0)$ diverges. Carrying the low-frequency weight through the Ramsey phase integral gives a decay that is Gaussian in time, not exponential, with only a weak logarithmic dependence on the averaging time.

where: $A$ — $1/f$ noise amplitude, $S_\lambda(\omega)=A/\lvert\omega\rvert$; $\omega_{\rm ir}$ — infrared cutoff set by the total measurement time. Reading: the Gaussian envelope is the fingerprint of $1/f$ dephasing, and its refocusability by echo is what makes dynamical decoupling effective. Two mechanisms dominate real superconducting devices. The first is Purcell decay: in the dispersive regime the qubit inherits a small photonic admixture from its lossy readout resonator and decays at the cavity rate weighted by the square of the coupling-to-detuning ratio.

where: $\kappa$ — photon decay rate of the readout cavity; $g$ — qubit–cavity coupling; $\Delta=\omega_{01}-\omega_r$ — qubit–cavity detuning. Reading: the $1/\Delta^2$ scaling means detuning sharply lengthens the Purcell-limited $T_1$, at the cost of a weaker dispersive readout — the tension a Purcell filter relieves. The second is quasiparticle tunneling across the Josephson junction, which even at a density as low as $x_{qp}\sim 10^{-6}$ can dominate $T_1$ in an otherwise clean device.

where: $x_{qp}$ — density of non-equilibrium quasiparticles normalized to the Cooper-pair density; $\Delta_{sc}$ — superconducting gap. Reading: the rate is linear in $x_{qp}$ with a kinematic factor from the BCS density of states near the gap edge, so quasiparticle poisoning is fought by gap engineering and shielding, not pulse tricks.

The interplay of $T_1$ and $T_2$ has a handful of clean regimes that between them cover almost every real qubit.

1 · Extracting the pure-dephasing time (confidence builder)

A superconducting qubit is measured to have $T_1 = 80\ \mu\text{s}$ and, in a Hahn-echo experiment, $T_2 = 50\ \mu\text{s}$. Find the pure-dephasing time $T_\phi$ and check that the numbers are self-consistent.

Step 1. Rearrange the master relation (2) for $T_\phi$:

Step 2. Invert: $T_\phi = 8000/110 \approx \mathbf{72.7\ \mu s}$. The dephasing time is comparable to $T_1$, so this qubit is neither purely relaxation-limited nor catastrophically dephasing-dominated.

Step 3. Consistency check: the ceiling is $2T_1 = 160\ \mu\text{s}$, and the measured $T_2 = 50\ \mu\text{s}$ sits comfortably below it, so the result is physical. $T_\phi \approx 72.7\ \mu s$, and $T_2 = 50 < 2T_1 = 160\ \mu s$, so the measurement is self-consistent.

2 · Purcell-limited relaxation and the value of detuning

A transmon couples to its readout resonator with $g/2\pi = 100\ \text{MHz}$, the cavity linewidth is $\kappa/2\pi = 1\ \text{MHz}$, and the qubit sits $\Delta/2\pi = 1\ \text{GHz}$ below the cavity. What Purcell-limited $T_1$ does this imply, and how does it change if the detuning is doubled to $2\ \text{GHz}$?

Step 1. The coupling-to-detuning ratio is $g/\Delta = 100\ \text{MHz}/1\ \text{GHz} = 0.1$, so $(g/\Delta)^2 = 0.01$. The $2\pi$ factors cancel because both are frequencies.

Step 2. Apply equation (6): $\Gamma_{\rm Purcell}=\kappa\,(g/\Delta)^2 = (2\pi\times 10^6\ \text{s}^{-1})(0.01) = 6.28\times 10^4\ \text{s}^{-1}$. Invert to get the Purcell-limited relaxation time $T_1^{\rm Purcell}=1/\Gamma_{\rm Purcell} \approx \mathbf{15.9\ \mu s}$.

Step 3. Because $\Gamma_{\rm Purcell}\propto 1/\Delta^2$, doubling the detuning cuts the rate by four and quadruples the time: $T_1^{\rm Purcell}\to \mathbf{63.7\ \mu s}$. Detuning from 1 to 2 GHz lifts the Purcell-limited $T_1$ from 15.9 to 63.7 μs, at the price of a four-times weaker dispersive shift and slower readout — the trade-off a Purcell filter removes.

3 · A coherence budget for gates

A qubit has $T_2 = 100\ \mu\text{s}$ and single-qubit gates take $t_g = 25\ \text{ns}$. Roughly how many coherent operations fit inside a coherence time, and what error per gate does that imply?

Step 1. The number of gates before coherence is lost is $N \approx T_2/t_g = (100\times 10^{-6})/(25\times 10^{-9}) = \mathbf{4000}$.

Step 2. The corresponding decoherence error accrued during one gate is $\varepsilon \approx t_g/T_2 = 25\ \text{ns}/100\ \mu\text{s} = \mathbf{2.5\times 10^{-4}}$.

Step 3. This is the quantity the DiVincenzo criteria demand be small, and it sits below the surface-code fault-tolerance threshold (a few $\times 10^{-3}$). About 4000 coherent operations per $T_2$, a decoherence floor of $2.5\times 10^{-4}$ per gate — which is why longer $T_2$ translates directly into deeper circuits.

Decoherence is one of those ideas that was implicit in quantum mechanics for decades before anyone named it. The relaxation times came first, from magnetic resonance; the master equations came next, from the theory of open systems; and the recognition that entanglement with an environment is what turns quantum into classical — and what limits every qubit — was the work of a small group of theorists who insisted on taking the environment seriously rather than wishing it away.

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, Ch. 8 (quantum operations and noise) · Preskill, Lecture Notes on Quantum Computation, Ch. 3 (open systems and decoherence) · Watrous, The Theory of Quantum Information, Ch. 2 (channels and CPTP maps) · Wilde, Quantum Information Theory, Ch. 4 (quantum channels) · Breuer & Petruccione, The Theory of Open Quantum Systems, Ch. 3 (Markovian master equations) · Krantz et al., "A Quantum Engineer's Guide to Superconducting Qubits", Appl. Phys. Rev. 6, 021318 (2019) · Zurek, "Decoherence, einselection, and the quantum origins of the classical", Rev. Mod. Phys. 75, 715 (2003). Key results: Worked Example 1 gives $T_\phi\approx 72.7\ \mu$s with $T_2<2T_1$; Example 2 gives $T_1^{\rm Purcell}=15.9\to 63.7\ \mu$s under $\Delta\to 2\Delta$; Example 3 gives $N\approx 4000$ gates and $\varepsilon\approx 2.5\times10^{-4}$; the Ramsey and echo filter scalings go as $\propto t^2/4$ and $\propto(\omega t)^2$ respectively. No platform-specific record coherence numbers are quoted, as state-of-the-art values evolve.

Prerequisites. DiVincenzo Criteria · Density matrices and open systems · Bloch sphere and qubit dynamics. Leads to. The Transmon Qubit · Dispersive Readout · Sideband Cooling · Quantum error correction.

Ask the tutor: Give me your qubit's $T_1$, $T_2$, and pulse sequence, and I will tell you which bath is limiting you and whether echo, a sweet spot, or a Purcell filter is your best next move.