Decoherence and Quantum Noise

A qubit's state is fully quantum only as long as it is isolated. The moment it interacts with any external degree of freedom — a phonon in the substrate, a stray photon, a fluctuating magnetic field, a nearby two-level defect — the qubit and that degree of freedom become entangled. If we cannot track the environment (and we never can, in practice), the correct description of the qubit alone is obtained by tracing out the environment from the joint state. This produces a reduced density matrix $\rho$ that no longer evolves under a single unitary. Its evolution is instead governed by a completely positive, trace-preserving (CPTP) map — a quantum channel. Decoherence is the general name for this loss of the qubit's unitary, coherent character due to entanglement with an inaccessible environment. Formally, suppose the qubit (system $S$) and environment ($E$) start in a product state $\rho_S \otimes \rho_E$ and evolve together under a joint unitary $U_{SE}$ generated by the total Hamiltonian $H = H_S + H_E + H_{int}$. The reduced state of the qubit after time $t$ is

Even though $U_{SE}$ is unitary on the joint system, the partial trace over $E$ makes the map on $\rho_S$ alone non-unitary in general. This is the mathematical seed of everything that follows: irreversibility for the qubit is not fundamental, it is the price of ignorance about the environment. Decoherence splits into two physically and mathematically distinct channels, both of which act on the off-diagonal and diagonal elements of $\rho$ differently:

• Relaxation (energy/amplitude damping, timescale T1): the qubit exchanges energy with the environment. An excited qubit ($|1\rangle$) spontaneously emits energy (a photon, a phonon) and decays to $|0\rangle$, analogous to spontaneous emission in atomic physics. T1 governs how fast the populations — the diagonal entries $\rho_{00}, \rho_{11}$ — relax to their thermal equilibrium values.
• Dephasing (timescale T2, or "pure dephasing" T2\, or $T_\phi$):</strong> the qubit's relative phase* between $|0\rangle$ and $|1\rangle$ is randomized by environmental fluctuations that couple to the qubit's energy without causing transitions. This destroys the off-diagonal coherences $\rho_{01}, \rho_{10}$ without necessarily changing the populations at all.

"Complexity" here means physical/temporal cost, not algorithmic time complexity — but the reasoning is exactly analogous to a runtime analysis: we derive a decay law and justify it from the underlying generator (the Lindblad equation), then quantify how quickly information is lost per unit time and per gate.

Population decay (T1) — derivation and justification. From the Lindblad master equation with jump operator $\sigma_-$ and rate $\gamma = 1/T_1$, taking the $\langle 1|\rho|1\rangle$ matrix element gives the ODE

This is the same structure as radioactive decay or spontaneous emission: a constant hazard rate $\gamma$ per unit time, giving the memoryless exponential law. The correctness of the model rests on the Markov approximation: the bath correlation time must be much shorter than $T_1$, so that at every instant the qubit "forgets" its history and decays at a rate depending only on its current state. This is justified physically when the bath spectral density is broad and smooth near the qubit frequency $\omega_0$ (a "flat" bath, e.g., a large photon or phonon continuum).

Coherence decay (T2) — derivation and justification. The off-diagonal element obeys

The factor $\tfrac12\gamma$ (not $\gamma$) appears because coherence involves one excitation's worth of amplitude decay (a factor $e^{-\gamma t/2}$ in the wavefunction picture, squared conceptually but here appearing linearly in the density matrix element which is bilinear in amplitudes — each of $|0\rangle,|1\rangle$ contributes, but only the excited component decays). This is why relaxation alone limits $T_2 \le 2T_1$: even with zero extra pure dephasing ($\gamma_\phi = 0$), coherence still decays at half the population rate translated through the factor of 2, giving $T_2 = 2T_1$ as the best case.

Why the inequality $T_2 \le 2T_1$ must hold — a clean argument. Coherence $\rho_{01}$ requires both basis states to remain well-defined and connected in phase. If the excited-state population decays at rate $1/T_1$, some coherence is destroyed automatically, contributing a decoherence rate of $1/(2T_1)$ to the off-diagonal element (half the population rate, since amplitude decay enters the density matrix element as one factor of the amplitude decay, not two). Any additional environmental process that randomizes phase without transitions (pure dephasing, rate $1/T_\phi \ge 0$) can only add to this rate, never subtract from it, because decoherence rates from independent Lindblad channels add. Hence $1/T_2 \ge 1/(2T_1)$, i.e., $T_2 \le 2T_1$, with equality iff $T_\phi \to \infty$ (no extra pure dephasing).

Tradeoffs and honesty about the model. The Lindblad/Markov picture is an approximation, not an exact description:

• It assumes weak coupling and a short bath correlation time. Real environments (1/f charge noise, flux noise) are often non-Markovian — slowly fluctuating, not memoryless — which is why the naive Ramsey-derived $T_2^*$ (sensitive to slow drift) differs from the echo-derived $T_2^{echo}$ (which cancels slow drift via a refocusing pulse) and from the "intrinsic" $T_2$ in the Lindblad sense.
• The formula $1/T_2 = 1/(2T_1) + 1/T_\phi$ assumes the two noise channels are statistically independent and both individually Markovian; in practice cross-correlations and 1/f noise spectra complicate the simple additive picture, and reported $T_2$ values are always sequence-dependent (Ramsey vs. echo vs. CPMG) unless stated otherwise.
• The exponential decay laws are the ensemble-averaged prediction; a single-shot experiment shows stochastic quantum jumps, and the exponential emerges only after averaging over many runs — a subtlety that matters when thinking about quantum trajectories and continuous monitoring.

Consider a superconducting transmon qubit with measured $T_1 = 50\ \mu s$ and pure dephasing time $T_\phi = 40\ \mu s$ (typical order-of-magnitude values for a good 2020s-era device). We trace both processes acting together on a qubit initialized in the equal superposition state $|\psi(0)\rangle = \tfrac{1}{\sqrt2}(|0\rangle + |1\rangle)$, i.e.

Step 1 — compute the effective $T_2$.

Note this correctly satisfies $T_2 \le 2T_1 = 100\ \mu s$, with the pure dephasing dominating (0.025 vs 0.010), consistent with $T_2 \ll 2T_1$ being the typical experimental regime.

Step 2 — evolve the populations (T1 channel only affects diagonal). At time $t$, with no re-excitation (zero-temperature bath, the standard idealization):

Step 3 — evolve the coherence (T2 governs off-diagonal magnitude, plus free precession phase $\omega_0 t$ which we set aside by working in the rotating frame, so the amplitude alone matters here).

Step 4 — tabulate at several times (in $\mu s$), dropping the oscillatory rotating-frame phase:

(Purity computed as $\rho_{00}^2 + \rho_{11}^2 + 2|\rho_{01}|^2$.)

Reading the table. At $t=0$ the state is pure ($\mathrm{Tr}\,\rho^2 = 1$), a genuine coherent superposition. By $t = 20\ \mu s$ — well under $T_1$ — the coherence $|\rho_{01}|$ has already dropped by nearly half, because $T_2 \approx 28.6\ \mu s$ is much shorter than $T_1 = 50\ \mu s$: dephasing destroys usable quantum information long before the qubit visibly relaxes. By $t = 100\ \mu s$ ($2T_1$), the coherence is essentially gone ($|\rho_{01}| \approx 0.009$) while the populations have relaxed most of the way to the ground-state equilibrium ($\rho_{11} \to 0$ at $T=0$ bath temperature). This numerically confirms the qualitative claim from the theory section: dephasing is the faster process and the tighter constraint on how long a superposition survives, which is exactly why gate-error budgets in error correction are dominated by $T_2$, not $T_1$, in dephasing-limited hardware.

The theory of decoherence grew out of the broader program of understanding open quantum systems and quantum measurement, connecting statistical mechanics, quantum optics, and quantum information.

Every quantum hardware platform reports T1 and T2 as headline specifications, because they set the "clock speed" available before information is lost: IBM and Google superconducting processors, IonQ and Quantinuum trapped-ion systems, and photonic platforms from Xanadu and PsiQuantum all publish these numbers as the primary benchmark of qubit quality. Gate fidelities, algorithm depth budgets, and the entire feasibility case for near-term ("NISQ") quantum computing are built directly on T1/T2 numbers, since a circuit whose runtime exceeds T2 by more than a small factor produces essentially classical noise instead of a quantum answer. Quantum error correction, the subject the next layer of theory addresses, exists specifically because T1 and T2 are finite and unavoidable at the physical level: no amount of engineering eliminates coupling to the environment entirely, so computation must be encoded redundantly across many physical qubits to detect and correct the errors that decoherence introduces faster than any single qubit can be improved. Understanding the T1/T2 split also directly informs which error channel — amplitude damping, phase damping, or a combination like the depolarizing channel — best approximates a given piece of hardware, which in turn determines which error-correcting codes and decoders are optimal for that device. Materials science efforts (better superconducting substrates, purer silicon, magnetically shielded ion traps) are direct engineering responses aimed at pushing T1 and T2 higher, and dynamical-decoupling pulse sequences (spin echoes, CPMG) are software-level mitigations that extend effective coherence time without changing the underlying hardware.

Ask the tutor: How does the continuous T1/T2 decoherence picture get converted into the discrete Pauli error channels used in quantum error correction?