Crosstalk

Crosstalk is the failure of a quantum operation to mind its own business. You ask the control system to act on one qubit — or to run two gates side by side — and you tacitly assume the rest of the register, and all of its correlations, are left as they were. Ideal multiqubit operations encode that assumption as a tensor product: a gate on qubit $A$ is $U_A\otimes I_{\bar A}$ , and two simultaneous gates factor as $U_A\otimes U_B$ . Real hardware honours neither. The generator of the dynamics carries coupling terms that entangle $A$ with its neighbours even when nothing is driven, and the control fields meant for $A$ spill onto $B$ . What comes out is a joint channel that refuses to factorize and whose errors are correlated in space and time. This article is the deep-tier synthesis of the noise thread. It takes the reduced-dynamics picture from aq-qci-open-quantum-systems-reduced-dynamics, the operator-sum language of aq-qci-kraus-representation-q9, the complete-positivity constraints of aq-qci-cptp-maps-quantum-operations, and the continuous-time generators of aq-qci-lindblad-gksl-master-equation, and asks one question: when a multiqubit map fails to factor, what exactly is the non-factoring part, and how big is it? Three consequences organize everything below. First, the dominant coherent crosstalk in fixed-frequency superconducting processors is an always-on $Z\otimes Z$ interaction — a residual conditional phase that accrues during every idle and every gate, born from the anharmonicity of coupled transmons. Second, that interaction is genuinely two-body: it turns product states into entangled ones, so no single-qubit recalibration can undo it. Third, crosstalk violates the very assumption error-correction thresholds are built on — locality and independence — so its correlated structure is a threat to the fault-tolerance argument, not a small correction. Tunable couplers, echoes, frequency allocation, and simultaneous randomized benchmarking are the working responses.

Key takeaways

Crosstalk is any deviation of a multiqubit operation from the ideal tensor product of independent local gates — the non-factorizing part of the realized channel.
Its three physical origins are static qubit–qubit coupling (quantum crosstalk, a residual $ZZ$ ), imperfectly local control fields (classical crosstalk), and spectator effects that make an active gate depend on an idle qubit's state.
The leading coherent term is an always-on $\tfrac{\zeta_{ZZ}}{4}Z_1Z_2$ ; with this normalization the conditional phase accrued in time $t$ is exactly $\phi=\zeta_{ZZ}t$ , and $\phi=\pi$ is a controlled- $Z$ — an idle can silently entangle.
In fixed-frequency transmons $\zeta_{ZZ}=\dfrac{2g^2(\alpha_1+\alpha_2)}{(\Delta+\alpha_1)(\Delta-\alpha_2)}$ : the residual $ZZ$ vanishes for harmonic oscillators, so it is fundamentally an anharmonicity effect.
Coherent crosstalk is a non-factoring unitary (reversible by echoes and couplers); incoherent crosstalk is correlated Kraus operators $K_{ab}\neq A_a\otimes B_b$ that a fixed unitary cannot undo — quantified by the diamond norm and detected as $r_{\mathrm{sim}}>r_{\mathrm{iso}}$ in simultaneous randomized benchmarking.
Key idea: crosstalk is precisely the non-factorizing part of a multiqubit operation — write $\Lambda=(\bigotimes_i\Lambda_i)\circ\Lambda_{\mathrm{xtalk}}$ ; the dominant coherent piece is generated by $\sum_{i<j}\tfrac{\zeta_{ij}}{4}Z_iZ_j$ plus off-resonant drive terms, and everything that does not commute with the ideal factorized gate is error.

The core insight

Crosstalk is precisely the non-factorizing part of a multiqubit operation: write the realized map as $\Lambda=(\bigotimes_i\Lambda_i)\circ\Lambda_{\mathrm{xtalk}}$ ; the dominant coherent piece is generated by an effective $\sum_{i<j}\tfrac{\zeta_{ij}}{4}Z_iZ_j$ term plus off-resonant drive terms, and everything that does not commute with the ideal factorized gate is error.

Where this lives today

Crosstalk is the error that scaling made unignorable. On a two-qubit demo it is a nuisance; on a hundred-qubit lattice it is a design constraint that shapes the whole processor. The dominant response in transmon arrays is the tunable coupler — a frequency-adjustable element placed between qubits so its two virtual exchange paths interfere destructively and drive $\zeta_{ZZ}\to0$ during idles, restoring channel factorization; such couplers are now standard in the large processors from Google and IBM. Where hardware cannot null the coupling, dynamical decoupling does it in software: because the $ZZ$ generator commutes with $Z_1Z_2$ , an $X$ pulse inserted halfway through an idle flips one qubit's $Z$ , reverses the interaction, and refocuses the accumulated phase to zero — the spin-echo trick, seventy years on. Fixed-frequency architectures such as IBM's heavy-hex layout add frequency allocation: a constraint solver detunes neighbours enough to keep off-resonant driving and resonant exchange small while dodging frequency collisions, and crosstalk-aware compilers avoid simultaneous gates on strongly coupled pairs. Characterization is its own discipline: simultaneous randomized benchmarking reports the gap $r_{\mathrm{sim}}-r_{\mathrm{iso}}$ that flags addressability errors, and Hamiltonian tomography maps the full $\zeta_{ij}$ table used to route circuits. The open frontier is fault tolerance itself — because crosstalk produces spatially correlated, non-Pauli, sometimes non-Markovian errors, it violates the independence assumptions of the cleanest threshold theorems, and how much correlated error a surface code can absorb before its distance scaling breaks is an active research question.

The physical picture

Picture two pendulums joined by a weak spring. Push one and the other stirs; push neither and the spring still trades phase back and forth between them. The spring is the always-on qubit–qubit coupling, and the slow phase transfer is the residual $ZZ$ interaction — coherent crosstalk you can in principle reverse but never simply ignore. Now bundle the two drive wires together. A signal meant for the first pendulum induces a little voltage on the second wire; that leaked drive is classical control crosstalk, a property of the apparatus rather than of the qubits, and it shrinks as the control hardware becomes more isolated and more frequency-selective. The two mechanisms are drawn together in the first figure, with the number a device team would actually quote.

Figure

Two coupled transmons: an always-on Z⊗Z of rate ζ_ZZ ≈ 2π × 120 kHz set by exchange g and detuning Δ, plus microwave leakage c·Ω from the drive on Q1 onto Q2 — the quantum and classical origins of crosstalk, both of which stop the joint map from factorizing

The crucial conceptual point is factorization. An ideal multiqubit operation is a tensor product; crosstalk is exactly the part that refuses to be one. Anything that makes $\Lambda\neq\bigotimes_i\Lambda_i$ counts, whether coherent — a Hamiltonian cross term such as $Z_1Z_2$ — or incoherent — correlated Kraus operators a stochastic bath imposes on two qubits at once. The coherent piece is the more insidious, because it accrues silently: no gate, no photon, no line in the schedule, only time passing while an always-on $ZZ$ turns the crank. The second figure makes that concrete for a representative $120\ \mathrm{kHz}$ coupling: the conditional phase a pair of idle qubits accumulates, and the spurious entanglement it creates from a perfectly separable input.

Figure

Conditional phase and spurious entanglement from an always-on ZZ of f_ZZ = 120 kHz: a product state left to idle accrues φ = 2π f_ZZ t, so after 1 μs φ = 0.75 rad ≈ 43° and an initially separable product state has gained concurrence C = sin(φ/2) = 0.37; by φ = π at t = 4.17 μs the idle alone has enacted a maximally entangling controlled-Z

The core equations

Everything starts from one definition: crosstalk is what is left when you divide the realized channel by the ideal factorized one. Writing the realized multiqubit superoperator as the ideal tensor product followed by a residual map isolates that leftover exactly.

\Lambda \;=\; \Big(\bigotimes_i \Lambda_i\Big)\circ \Lambda_{\mathrm{xtalk}}, \qquad \Lambda_{\mathrm{xtalk}}=\mathcal{I}\ \Leftrightarrow\ \text{no crosstalk} \qquad (1)

where: $\Lambda$ — realized multiqubit CPTP map of an intended operation; $\Lambda_i$ — ideal local map on qubit $i$ ; $\Lambda_{\mathrm{xtalk}}$ — residual non-local map carrying all crosstalk, the identity exactly when the operation factorizes. Reading it: crosstalk is a statement about a ratio of channels, not about one qubit's error rate. The coherent part of $\Lambda_{\mathrm{xtalk}}$ comes from the Hamiltonian terms that couple different qubits. For superconducting hardware the effective Hamiltonian, in the qubit subspace and rotating frame, collects a single-qubit part, a two-body $ZZ$ part, and the drives.

H \;=\; \sum_i \frac{\omega_i}{2} Z_i \;+\; \sum_{i<j} \frac{\zeta_{ij}}{4}\,Z_iZ_j \;+\; \sum_i \frac{\Omega_i(t)}{2}\,(\cos\phi_i\,X_i + \sin\phi_i\,Y_i) \qquad (2)

where: $\omega_i$ — qubit transition frequency; $\zeta_{ij}$ — residual $ZZ$ rate between qubits $i,j$ (rad/s); $\Omega_i(t)$ — drive (Rabi) amplitude; $\phi_i$ — drive phase setting the rotation axis. The $\tfrac14$ normalization makes $\zeta_{ij}=E_{11}-E_{10}-E_{01}+E_{00}$ the directly measured conditional-phase rate; the ideal circuit keeps only $\tfrac{\omega_i}{2}Z_i$ and the intended drive. The $ZZ$ term is not put in by hand. It is the residue of the exchange coupling between two anharmonic oscillators once the fast excitation swap is adiabatically eliminated; its size is set by the coupling, the detuning, and — decisively — the anharmonicity.

\zeta_{ZZ} \;=\; 2g^2\!\left(\frac{1}{\Delta-\alpha_2}-\frac{1}{\Delta+\alpha_1}\right) \;=\; \frac{2g^2(\alpha_1+\alpha_2)}{(\Delta+\alpha_1)(\Delta-\alpha_2)} \qquad (3)

where: $g$ — exchange coupling; $\Delta=\omega_1-\omega_2$ — qubit–qubit detuning; $\alpha_k<0$ — transmon anharmonicity. Reading it: for harmonic modes $\alpha_k=0$ the two terms cancel and $\zeta_{ZZ}=0$ , so the static $ZZ$ of transmons is an anharmonicity effect; it is derived in full below. Restrict the dynamics to the $ZZ$ term and exponentiate, and the coherent crosstalk becomes a diagonal, entangling unitary. The single physically meaningful phase is the conditional one — the combination that cannot be absorbed into single-qubit $Z$ rotations.

U_{ZZ}(t)=e^{-i\frac{\zeta_{ZZ}}{4}t\,Z_1Z_2}=\mathrm{diag}\!\big(e^{-i\zeta t/4},e^{+i\zeta t/4},e^{+i\zeta t/4},e^{-i\zeta t/4}\big),\qquad \phi=\zeta_{ZZ}\,t \qquad (4)

where: $\phi$ — conditional phase, $\phi=\arg\!\big(\tfrac{\langle11|U|11\rangle\,\langle00|U|00\rangle}{\langle01|U|01\rangle\,\langle10|U|10\rangle}\big)^{-1}=\zeta_{ZZ}t$ ; $t$ — idle or gate duration. At $\phi=\pi$ this unitary is a controlled- $Z$ : an uncorrected idle enacts a two-qubit gate. Classical crosstalk enters through the drive terms. A pulse meant for a neighbour reaches a spectator off-resonantly, detuned by $\delta$ , and produces a second-order AC Stark shift — a state-dependent phase that scales as the drive power over the detuning.

\Omega_{\mathrm{eff}} \;=\; \frac{\Omega^2}{2\,\delta}, \qquad \delta=\omega_{\mathrm{spec}}-\omega_{\mathrm{drive}} \qquad (5)

where: $\Omega$ — Rabi amplitude of the leaked drive on the spectator; $\delta$ — detuning of the drive from the spectator's transition; $\Omega_{\mathrm{eff}}$ — effective rate of unintended evolution. It vanishes as $\delta\to\infty$ , which is why frequency allocation is a crosstalk-mitigation tool. Not all crosstalk is coherent. When neighbouring qubits share a fluctuating bath, the joint channel has Kraus operators that do not factor — the signature of correlated, incoherent crosstalk that no fixed unitary can reverse.

\mathcal{E}(\rho)=\sum_{a,b}K_{ab}\,\rho\,K_{ab}^\dagger,\qquad K_{ab}\neq A_a\otimes B_b\ \Rightarrow\ \text{correlated crosstalk} \qquad (6)

where: $K_{ab}$ — Kraus operators of the realized joint channel; $A_a\otimes B_b$ — the product form a factorizing (independent-noise) channel would have. If every $K_{ab}$ factored, the channel would be $\mathcal{E}_A\otimes\mathcal{E}_B$ and there would be no crosstalk. Finally, to compare devices and to feed fault-tolerance arguments one needs a single operational number for "how much crosstalk". The diamond norm of the difference between the realized and factorized channels supplies the worst-case, composable measure.

\epsilon_{\mathrm{xtalk}} \;=\; \tfrac12\,\Big\|\,\Lambda-\textstyle\bigotimes_i\Lambda_i\,\Big\|_{\diamond} \qquad (7)

where: $\|\cdot\|_\diamond$ — diamond (completely bounded trace) norm, the largest distinguishability of two channels over all inputs entangled with a reference; $\epsilon_{\mathrm{xtalk}}$ — operationally meaningful size of crosstalk. Its experimental shadow is the gap $r_{\mathrm{sim}}-r_{\mathrm{iso}}>0$ between simultaneous and isolated randomized-benchmarking error rates.

Derivation AResidual ZZ from dispersive transmon coupling

The goal is to earn the always-on conditional-phase coupling of equation (3) rather than assert it — to show that two capacitively coupled transmons develop a static $Z\otimes Z$ interaction the moment their anharmonicity is nonzero, with rate $2g^2(\alpha_1+\alpha_2)/[(\Delta+\alpha_1)(\Delta-\alpha_2)]$ . The strategy is second-order perturbation theory in the exchange coupling: let the doubly-excited state $|11\rangle$ borrow energy from the anharmonic levels $|02\rangle,|20\rangle$ , subtract the single-excitation shifts, and keep the genuinely two-body remainder.

Step 1. Write the coupled Hamiltonian $H=\sum_k(\omega_k a_k^\dagger a_k+\tfrac{\alpha_k}{2}a_k^\dagger a_k^\dagger a_k a_k)+g(a_1^\dagger a_2+a_1 a_2^\dagger)$ and truncate each transmon to $\{|0\rangle,|1\rangle,|2\rangle\}$ , so a single mode has energies $E_{|0\rangle}=0$ , $E_{|1\rangle}=\omega$ , $E_{|2\rangle}=2\omega+\alpha$ . Why: the anharmonicity $\alpha$ is what makes $|2\rangle$ sit at $2\omega+\alpha$ rather than $2\omega$ , and that displacement is the entire source of the effect — a harmonic ladder would have $\alpha=0$ .

Step 2. Treat $g$ as the perturbation and find the states it connects to $|11\rangle$ . The exchange operator moves one quantum, so $|11\rangle$ couples to $|02\rangle$ and $|20\rangle$ with matrix elements $\langle02|g\,a_1a_2^\dagger|11\rangle=g\sqrt2$ and $\langle20|g\,a_1^\dagger a_2|11\rangle=g\sqrt2$ , the $\sqrt2$ coming from $a^\dagger|1\rangle=\sqrt2\,|2\rangle$ . Why: the bosonic enhancement $\sqrt2$ is exactly what makes the $|11\rangle$ shift differ from twice a single-excitation shift — it is the algebraic fingerprint of the second excitation.

Step 3. Apply the second-order formula $\delta E=\sum_m|\langle m|V|11\rangle|^2/(E_{11}-E_m)$ with unperturbed energies $E_{11}=\omega_1+\omega_2$ , $E_{02}=2\omega_2+\alpha_2$ , $E_{20}=2\omega_1+\alpha_1$ . The denominators are $E_{11}-E_{02}=\Delta-\alpha_2$ and $E_{11}-E_{20}=-\Delta-\alpha_1$ , giving $\delta E_{11}=\dfrac{2g^2}{\Delta-\alpha_2}-\dfrac{2g^2}{\Delta+\alpha_1}$ . Why: writing the denominators as $\Delta\mp\alpha$ shows the anharmonicity sitting right next to the detuning — it is the reason the two virtual paths do not cancel.

Step 4. Subtract the single-excitation shifts. The states $|10\rangle$ and $|01\rangle$ couple only to each other, with shifts $\delta E_{10}=g^2/\Delta$ and $\delta E_{01}=-g^2/\Delta$ , which cancel in the combination $\zeta_{ZZ}=\delta E_{11}-\delta E_{10}-\delta E_{01}+\delta E_{00}$ since $\delta E_{00}=0$ . What survives is $\zeta_{ZZ}=2g^2\big(\tfrac{1}{\Delta-\alpha_2}-\tfrac{1}{\Delta+\alpha_1}\big)=\dfrac{2g^2(\alpha_1+\alpha_2)}{(\Delta+\alpha_1)(\Delta-\alpha_2)}$ . Why: the $E_{11}-E_{10}-E_{01}+E_{00}$ combination is designed to kill everything that is a sum of one-body pieces, leaving only the irreducibly two-body energy — which is the $ZZ$ rate. *Three quick checks: setting $\alpha_1=\alpha_2=0$ gives $\zeta_{ZZ}=0$ , confirming the harmonic limit; the formula is dimensionally consistent, returning units of frequency; and with $g/2\pi=3$ MHz, $\Delta/2\pi=100$ MHz, $\alpha/2\pi=-330$ MHz it evaluates to $\zeta_{ZZ}/2\pi=+120$ kHz.*

Go deeper · Derivation B · 2 minCoherent crosstalk breaks tensor-product factorization

Derivation BCoherent crosstalk breaks tensor-product factorization

The aim is to prove that any nonzero coherent cross term makes the joint channel irreducibly non-local — that $U_{ZZ}$ cannot be written as $U_1\otimes U_2$ for any single-qubit factors — and to make it quantitative by computing the entanglement generated. The method is a clean contradiction: feed the map a product input and exhibit an output whose concurrence is nonzero, which a product unitary can never produce.

Step 1. Apply the $ZZ$ unitary of equation (4) to the product state $|{+}{+}\rangle=\tfrac12(|00\rangle+|01\rangle+|10\rangle+|11\rangle)$ . Because $U_{ZZ}$ is diagonal it only stamps phases: $|{+}{+}\rangle\mapsto\tfrac12\big(e^{-i\theta}|00\rangle+e^{i\theta}|01\rangle+e^{i\theta}|10\rangle+e^{-i\theta}|11\rangle\big)$ with $\theta=\zeta_{ZZ}t/4=\phi/4$ . Why: starting from a manifestly separable state is the whole trick — if the output turns out entangled, no product unitary could have produced it.

Step 2. Compute the concurrence of the output. For a two-qubit pure state with amplitudes $(c_{00},c_{01},c_{10},c_{11})$ the concurrence is $C=2|c_{00}c_{11}-c_{01}c_{10}|$ ; here $c_{00}=c_{11}=\tfrac12 e^{-i\theta}$ and $c_{01}=c_{10}=\tfrac12 e^{i\theta}$ , so $C=2\big|\tfrac14 e^{-2i\theta}-\tfrac14 e^{2i\theta}\big|=|\sin2\theta|=|\sin(\phi/2)|$ . Why: concurrence is the sharpest yes/no test for two-qubit entanglement, and getting a clean closed form $|\sin(\phi/2)|$ ties the entanglement directly to the conditional phase.

Step 3. Read off the contradiction and the physics. For any $\phi\notin\{0,2\pi,\dots\}$ the concurrence is strictly positive, so the output is entangled; but a product unitary $U_1\otimes U_2$ maps product states to product states and can never raise concurrence above zero. Therefore $U_{ZZ}\neq U_1\otimes U_2$ , and the crosstalk channel $\mathcal{E}_{ZZ}(\rho)=U_{ZZ}\rho\,U_{ZZ}^\dagger$ is irreducibly non-local. Why: this is the precise sense in which coherent crosstalk is not a single-qubit error — no local recalibration, on either qubit, can remove a genuinely two-body generator.

Step 4. Locate the special angles. At $\phi=\pi$ (so $\theta=\pi/4$ ) the concurrence saturates at $C=1$ : the idle interaction has become a maximally entangling controlled- $Z$ , up to local frames. At the representative $f_{ZZ}=120$ kHz, a $1\ \mu\mathrm{s}$ idle gives $\phi=0.754$ and $C=\sin(0.377)=0.368$ . Why: seeing that a mere waiting period reaches a full CZ at $\phi=\pi$ is the visceral statement of why coherent crosstalk cannot be tolerated in a deep circuit. *Three checks on the result: $C=|\sin(\phi/2)|$ gives $C(\phi=\pi)=1$ (a maximally entangling CZ) and $C(\phi=0.754)=0.368$ ; and because the map is unitary it preserves purity, so all lost fidelity is coherent and therefore refocusable by an echo.*

Go deeper · Derivation C · 1 minThe off-resonant AC Stark shift of a spectator

Derivation CThe off-resonant AC Stark shift of a spectator

The purpose is to ground the classical-crosstalk formula (5): to show that a drive intended for one qubit, arriving on a spectator detuned by $\delta$ , produces a coherent frequency shift $\Omega^2/(2\delta)$ and only a small, bounded population error. The route is to diagonalize the spectator's driven two-level Hamiltonian exactly and expand in the small ratio $\Omega/\delta$ — the regime a well-designed processor lives in.

Step 1. In the frame rotating at the drive frequency, the leaked drive presents the spectator with $H_{\mathrm{spec}}=\tfrac{\delta}{2}Z+\tfrac{\Omega}{2}X$ : a detuning term plus a transverse drive of Rabi rate $\Omega=c\,\Omega_{\mathrm{target}}$ set by the control-crosstalk ratio $c$ . Why: a pulse resonant on the target is off-resonant on the spectator by the qubit–qubit frequency difference, so the spectator precesses about a tilted axis rather than rotating cleanly.

Step 2. Diagonalize exactly: the eigenvalues are $\pm\tfrac12\sqrt{\delta^2+\Omega^2}$ , so the transition frequency shifts from $\delta$ to $\sqrt{\delta^2+\Omega^2}$ . Why: the driven splitting is the length of $(\Omega,0,\delta)$ , and its excess over $\delta$ is the entire Stark effect — still exact.

Step 3. Expand for $\Omega\ll\delta$ : $\sqrt{\delta^2+\Omega^2}\approx\delta+\dfrac{\Omega^2}{2\delta}$ , so the induced shift is $\Omega_{\mathrm{eff}}=\Omega^2/(2\delta)$ , a coherent, state-dependent phase rate, while the chance of actually flipping the spectator is bounded by $(\Omega/\delta)^2$ . Why: classical crosstalk shows up first as a phase error ( $\propto\Omega^2/\delta$ ), not a bit flip ( $\propto1/\delta^2$ ) — and both shrink as neighbours are detuned, the lever frequency allocation pulls. *As a numerical check: with $\Omega/2\pi=1.25$ MHz leakage and $\delta/2\pi=200$ MHz, the formula gives $\Omega_{\mathrm{eff}}/2\pi=3.9$ kHz and a maximum population leakage of $(\Omega/\delta)^2=3.9\times10^{-5}$ .*

Limiting cases

Crosstalk takes characteristic forms in well-separated physical regimes; the same coupling $g$ produces phase, population, or nothing depending on where the qubits sit.

Case	Result	Meaning
Dispersive, $\vert\Delta\vert\gg g$	$\zeta_{ZZ}\approx\dfrac{2g^2(\alpha_1+\alpha_2)}{(\Delta+\alpha_1)(\Delta-\alpha_2)}$	a static conditional phase accrues with no drive at all — the dominant coherent crosstalk of fixed-frequency transmons.
Resonant exchange, $\Delta\to0$	$H_{\mathrm{ex}}=g(\sigma^+_1\sigma^-_2+\sigma^-_1\sigma^+_2)$	on resonance the coupling swaps excitations (iSWAP-like): population, not just phase, leaks between qubits — strong coherent crosstalk.
Off-resonant drive	$\Omega_{\mathrm{eff}}=\dfrac{\Omega^2}{2\delta},\ P_{\mathrm{leak}}\approx(\Omega/\delta)^2$	a leaked drive detuned by $\delta$ gives an AC Stark phase and tiny rotation — classical crosstalk that vanishes as $\delta\to\infty$ .
Tunable coupler / echo	$\zeta_{ZZ}\to0,\ \Lambda\to\bigotimes_i\Lambda_i$	when a coupler nulls $ZZ$ or a refocusing pulse averages it away, the map factorizes and crosstalk is suppressed to leading order.
Harmonic modes	$\alpha_k=0\Rightarrow\zeta_{ZZ}=0$	the two virtual paths cancel: no static $ZZ$ without anharmonicity, the reason qubits, not oscillators, need this care.

Common trapSmall single-qubit errors mean two-qubit errors are independent

This is the most expensive misconception in the subject. Single-qubit benchmarks — a clean $T_1$ , a clean $T_2$ , a $99.9\%$ isolated Clifford fidelity — say nothing about whether the joint map factorizes. Crosstalk lives entirely in the correlations: an always-on $ZZ$ of a few tens of kHz can dominate the idle error budget while every qubit's isolated numbers look pristine, because it is a two-body effect that single-qubit characterization is blind to. You cannot recover the joint infidelity by multiplying single-qubit infidelities — the cross term is not in that product — and correlated errors can behave like higher-weight faults and quietly undercut a code's distance. The fix is to measure the thing that is actually correlated (simultaneous randomized benchmarking, Hamiltonian tomography of $\zeta_{ij}$ ), never to infer it from isolated numbers.

Worked examples

1 · An idle that entangles (fully worked)

Two fixed-frequency transmons have $g/2\pi=3\ \mathrm{MHz}$ , detuning $\Delta/2\pi=100\ \mathrm{MHz}$ , and equal anharmonicity $\alpha/2\pi=-330\ \mathrm{MHz}$ . They are prepared in $|{+}{+}\rangle$ and left idle for $t=1\ \mu\mathrm{s}$ with no gates scheduled. Find the residual $ZZ$ rate, the conditional phase, the entanglement generated, and the idle-state fidelity.

Step 1. Evaluate the $ZZ$ rate from equation (3). Numerator $2g^2(\alpha_1+\alpha_2)=2(3)^2(-660)=-11880\ \mathrm{MHz}^3$ ; denominator $(\Delta+\alpha_1)(\Delta-\alpha_2)=(100-330)(100+330)=(-230)(430)=-98900\ \mathrm{MHz}^2$ . The ratio is $\zeta_{ZZ}/2\pi=-11880/-98900=\mathbf{+0.120\ MHz}=120\ \mathrm{kHz}$ . The two negative factors make $\zeta_{ZZ}$ positive — a real, measurable coupling with no drive applied.

Step 2. Accumulate the conditional phase over the idle. With the normalization of equation (4), $\phi=\zeta_{ZZ}t=2\pi f_{ZZ}t=2\pi(120\times10^{3})(1\times10^{-6})=\mathbf{0.754\ rad}=43.2^\circ$ . This is the phase of $|11\rangle$ relative to what independent single-qubit evolution would predict — the part no $Z$ calibration can remove.

Step 3. Convert the phase to entanglement using the concurrence from Derivation 2, $C=|\sin(\phi/2)|=\sin(0.377)=\mathbf{0.368}$ . A state that began perfectly separable is now $37\%$ of the way to a Bell pair, purely from waiting. Nothing clicked, nothing was driven — the correlation is a gift of the always-on Hamiltonian.

Step 4. Compute the fidelity against the intended (identity) idle. The overlap is $\langle{+}{+}|U_{ZZ}|{+}{+}\rangle=\cos\theta$ with $\theta=\phi/4=0.1885$ , so $F=\cos^2(\phi/4)=\cos^2(0.1885)=\mathbf{0.965}$ . A single $1\ \mu\mathrm{s}$ idle costs $3.5\%$ fidelity to coherent crosstalk alone — larger than the entire decoherence budget of a good transmon over the same microsecond, and, because it is coherent, removable by an echo rather than by better $T_1$ .

2 · Control leakage onto a spectator

A single-qubit gate drives its target at Rabi rate $\Omega_{\mathrm{target}}/2\pi=25\ \mathrm{MHz}$ (a $20\ \mathrm{ns}$ $\pi$ -pulse). The control line couples a fraction $c=0.05$ of that amplitude onto a spectator detuned by $\delta/2\pi=200\ \mathrm{MHz}$ . Estimate the AC Stark phase and the leakage the spectator suffers during the gate.

Step 1. Find the leaked Rabi rate: $\Omega/2\pi=c\,\Omega_{\mathrm{target}}/2\pi=0.05\times25=\mathbf{1.25\ MHz}$ . This is the transverse drive the spectator actually sees.

Step 2. Apply equation (5): $\Omega_{\mathrm{eff}}/2\pi=\Omega^2/(2\delta)$ in ordinary-frequency units is $(1.25)^2/(2\times200)=1.5625/400=\mathbf{3.9\ kHz}$ . Over the $20\ \mathrm{ns}$ gate the spurious Stark phase is $\phi_{\mathrm{Stark}}=2\pi(3.9\times10^{3})(20\times10^{-9})=\mathbf{4.9\times10^{-4}\ rad}$ — a small, coherent, state-dependent error.

Step 3. Bound the population error: $P_{\mathrm{leak}}\approx(\Omega/\delta)^2=(1.25/200)^2=\mathbf{3.9\times10^{-5}}$ . Classical crosstalk here is orders of magnitude weaker than the coherent $ZZ$ of Example 1 — the general lesson that, in fixed-frequency hardware, the static $ZZ$ usually dominates the budget while control leakage is a smaller, phase-like correction that large $\delta$ suppresses further.

Physics of the portrait

Crosstalk is an old word wearing new physics. It was named a century ago by telephone engineers fighting the faint ghost of one conversation bleeding into another; its quantum incarnation — the residual $ZZ$ coupling that dominates coherent error in superconducting processors — is the same idea sharpened until a single unwanted phase can spoil a computation.

Origins · 1 minresidual ZZ coupling and how this was discovered

residual ZZ coupling

the always-on ζ Z₁Z₂ that dominates coherent crosstalk · dispersive circuit QED and NMR J-coupling

The residual $ZZ$ coupling is not a person but a phenomenon, and its lineage runs straight through twentieth-century physics. The word crosstalk was coined in telephony, where signal leaking between adjacent wire pairs was the limiting impairment of long-distance cables; the engineering instinct — isolate the lines, keep the channels independent — survives unchanged in quantum hardware. Its physical content, an always-on coupling between two-level systems, was understood decades before qubits by the nuclear-magnetic-resonance community: the scalar (or $J$ ) coupling between nuclear spins, explained by Gutowsky and co-workers in the early 1950s, is mathematically the same $Z\otimes Z$ interaction, and Erwin Hahn's spin echo (1950) is the very refocusing trick still used to cancel $ZZ$ in a circuit. When circuit quantum electrodynamics arrived in 2004 it brought the dispersive coupling of superconducting qubits into the microwave regime, and the transmon (2007) — a deliberately weakly anharmonic oscillator — made the second-order shift of equation (3) an unavoidable feature of any fixed-frequency pair. What NMR called $J$ -coupling, superconducting engineers now call the residual $ZZ$ rate. The concept is a bridge across three fields that each rediscovered the same enemy: an interaction that will not turn itself off.

History

1915

— crosstalk named as a limiting impairment in long-distance telephony

1950

— Hahn's spin echo refocuses unwanted couplings between spins

1953

— Gutowsky and co-workers explain scalar J-coupling between nuclear spins

2004

— circuit QED brings the dispersive coupling of qubits into the microwave regime

2012

— simultaneous randomized benchmarking is introduced to quantify addressability and crosstalk

Going further

Texts. Nielsen and Chuang, Quantum Computation and Quantum Information, Ch. 8 (quantum operations and error models) · Preskill, Lecture Notes on Quantum Computation, Ch. 3 (decoherence, master equations) and Ch. 7 (fault tolerance and correlated errors) · Watrous, The Theory of Quantum Information, Ch. 2–3 (channels, tensor structure, diamond norm) · Wilde, Quantum Information Theory, Ch. 4 (quantum channels) · Kitaev, Shen, and Vyalyi, Classical and Quantum Computation (error models). The residual- $ZZ$ formula $\zeta_{ZZ}=2g^2(\alpha_1+\alpha_2)/[(\Delta+\alpha_1)(\Delta-\alpha_2)]$ is derived by second-order perturbation theory with matrix elements $g\sqrt2$ to $|02\rangle,|20\rangle$ , and passes the harmonic-limit check $\alpha\to0\Rightarrow\zeta\to0$ as well as dimensional consistency. The normalization $H_{ZZ}=\tfrac{\zeta}{4}Z_1Z_2$ is chosen so that $\zeta_{ZZ}=E_{11}-E_{10}-E_{01}+E_{00}$ is the directly measured conditional-phase rate and $\phi=\zeta_{ZZ}t$ ; the concurrence formula $C=|\sin(\phi/2)|$ gives $C=1$ at $\phi=\pi$ (a CZ). The AC Stark shift $\Omega_{\mathrm{eff}}=\Omega^2/2\delta$ follows from exact diagonalization of $\tfrac{\delta}{2}Z+\tfrac{\Omega}{2}X$ and expansion in $\Omega/\delta$ . Worked-example numbers: $\zeta_{ZZ}/2\pi=120$ kHz, $\phi=0.754$ rad, $C=0.368$ , $F=0.965$ ; $\Omega_{\mathrm{eff}}/2\pi=3.9$ kHz, $P_{\mathrm{leak}}=3.9\times10^{-5}$ . The 120 kHz figure is a representative evaluation of equation (3) at the stated parameters, not a universal device value.

Prerequisites. Open Quantum Systems and Reduced Dynamics · Quantum Operations and CPTP Maps · Kraus Representation of Quantum Channels · Lindblad / GKSL Master Equation. Leads to. Pauli and Bit/Phase-Flip Channels · Markovian vs Non-Markovian Dynamics · Depolarizing Channel.

Ask the tutor: Ask me to turn your device's $g$ , $\Delta$ , and $\alpha$ into a $ZZ$ rate and idle error, to walk a product state through an always-on $ZZ$ step by step, or to contrast tunable couplers, echoes, and frequency allocation as crosstalk cures.

What is the difference between quantum and classical crosstalk?

They have different origins and different cures. Quantum (or static) crosstalk is a Hamiltonian effect: an always-present coupling between qubits — capacitive, inductive, or dipolar — that in the dispersive regime reduces to a residual $ZZ$ interaction. It accrues even when nothing is driven and is a genuine two-body term, so it entangles. Classical (or control) crosstalk is an apparatus effect: the field meant for one qubit reaches another through shared lines, electromagnetic leakage, and finite frequency selectivity, producing off-resonant driving and AC Stark shifts. Quantum crosstalk is fought with tunable couplers and echoes (it lives in the qubits); classical crosstalk with better isolation, pulse shaping, and frequency allocation (it lives in the wiring). In today's fixed-frequency transmons the static $ZZ$ usually dominates.

Why does the residual ZZ vanish for harmonic oscillators?

Because the effect is built entirely out of the anharmonic levels. In second-order perturbation theory the doubly-excited state $|11\rangle$ borrows energy through virtual transitions to $|02\rangle$ and $|20\rangle$ , and the energy denominators for those paths are $\Delta-\alpha_2$ and $-\Delta-\alpha_1$ . When you subtract the single-excitation shifts to isolate the genuinely two-body part, the result is proportional to $(\alpha_1+\alpha_2)$ . Set the anharmonicities to zero — a perfectly harmonic ladder — and the two virtual paths contribute equal and opposite shifts that cancel exactly, leaving $\zeta_{ZZ}=0$ . This is why linear resonators do not develop a static $ZZ$ and why transmons, which are deliberately made weakly anharmonic so their two lowest levels form a usable qubit, pay for that anharmonicity with an unavoidable coherent crosstalk.

How can an idle period entangle two qubits if I never apply a gate?

Because "idle" means "no gate in the schedule," not "no Hamiltonian." The always-on $ZZ$ term runs continuously, so during any waiting time the two-qubit state accumulates a conditional phase $\phi=\zeta_{ZZ}t$ . Since $Z_1Z_2$ is a genuine two-body operator, that phase turns a separable state into an entangled one — the concurrence grows as $|\sin(\phi/2)|$ — and at $\phi=\pi$ the idle has enacted a full controlled- $Z$ . Nothing was driven and no energy exchanged; the correlation is a coherent gift of the device Hamiltonian. This is exactly why circuits insert echo pulses during idle windows and why processors invest in tunable couplers: to switch off a coupling that would otherwise keep working while you wait.

How do I distinguish coherent from incoherent crosstalk?

By whether a fixed unitary can undo it. Coherent crosstalk is a unitary deviation generated by cross Hamiltonian terms such as $Z_1Z_2$ : it gives $\Lambda(\rho)=U\rho U^\dagger$ with a non-factorizing $U$ , it preserves the purity of states, and it is in principle fully reversible by an echo or a compensating gate. Incoherent crosstalk shows up as Kraus operators $K_{ab}$ that do not factor into $A_a\otimes B_b$ — correlated stochastic errors that a shared bath imprints on two qubits at once. These mix states, reduce purity, and cannot be reversed by any fixed unitary; they must be corrected by an error-correcting code, and their correlated structure is precisely what makes standard thresholds fragile. In practice the two are separated by process tomography or by comparing the coherent and incoherent parts of a benchmarked channel.

Why use the diamond norm instead of a single fidelity number?

Because crosstalk error depends on the input state and on entanglement with the rest of the register, and an average fidelity can hide the worst case. The diamond norm $\tfrac12\|\Lambda-\bigotimes_i\Lambda_i\|_\diamond$ measures the largest distinguishability between the realized and the ideal factorized channel over all inputs — including inputs entangled with a reference system the channel never touches. That worst-case quantity is what fault-tolerance proofs need, because it composes: the diamond distance of a sequence of operations is bounded by the sum of the individual distances. An average gate fidelity does not compose that way and can look large even when a specific, adversarially chosen input is corrupted badly. For threshold arguments, where the error is effectively worst-case, the diamond norm is the honest measure.

Does crosstalk actually threaten quantum error correction, or is it just extra noise?

It is a structural threat, not merely more noise. Threshold theorems that promise arbitrarily good logical qubits are typically proved assuming physical errors on different qubits are independent, or at most weakly correlated with a locality bound. Crosstalk breaks that assumption at the root: a static $ZZ$ or a shared control fault produces spatially correlated errors that act like higher-weight faults, and a code's protection scales with the weight of the errors it must catch. A distance- $d$ surface code corrects errors of weight up to $\lfloor(d-1)/2\rfloor$ acting independently, but a single correlated crosstalk event can flip several data qubits at once and defeat that scaling. This is why crosstalk mitigation is a prerequisite for fault tolerance rather than an afterthought, and why characterizing correlated error — not just average rate — is now standard in qualifying a processor.