Fluxonium Readout Techniques

• Fluxonium readout is a suite of measurement techniques that exploit dispersive coupling and flux tunability to distinguish the quantum state of a fluxonium qubit.
• It integrates methods like flux pulsing, shelving, fluorescence, and ancilla-mediated mapping to manage multilevel effects and non-QND channels.
• Implementations using circuit-QED architectures demonstrate high-fidelity, fast readout with detailed error accounting, paving the way for scalable quantum processors.

Fluxonium readout comprises the family of measurement techniques used to discriminate the quantum state of a fluxonium qubit, most commonly through dispersive coupling to a microwave mode, but also through shelving, fluorescence, ancilla-mediated mapping, and more speculative fluxon-scattering schemes. Its technical character is set by the conjunction of low qubit transition frequencies, large anharmonicity, strong multilevel effects, parity selection rules, and pronounced flux tunability. These features make fluxonium readout both unusually flexible and unusually sensitive to higher-level structure, so the subject is defined as much by control of non-QND channels as by raw signal separation (Bao et al., 2021, Bothara et al., 28 Jan 2025, Li et al., 19 Jul 2025, Watanabe et al., 22 Apr 2025).

1. Circuit-QED basis and flux dependence

A standard starting point is the fluxonium Hamiltonian

$$H_q(\Phi)=4E_C \hat n^2+\tfrac12 E_L \hat\varphi^2-E_J\cos(\hat\varphi-\Phi/\Phi_0\cdot\pi),$$

or closely related phase-basis forms differing only by convention for the external-flux term. The readout mode is typically a single resonator mode $a$, and in the dispersive regime the effective interaction is written

$$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$

For fluxonium, however, $\chi$ is intrinsically multilevel: a common approximation is

$$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$

with $\Delta_{ge}=\omega_q-\omega_r$ and $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$. The literature repeatedly emphasizes that the dominant contribution is often controlled by higher transitions or hybridization involving noncomputational states rather than by a simple two-level picture (Bao et al., 2021, Li et al., 19 Jul 2025, Stefanski et al., 2023).

This multilevel structure gives fluxonium a markedly nonuniform dispersive landscape. At the sweet spot, the qubit frequency is minimal and first-order insensitive to flux noise; in one device $f_{01}/2\pi\approx0.56$ GHz, and in another it is $\approx0.4$ GHz. Away from the sweet spot, both $f_{01}(\Phi)$ and higher transition frequencies move substantially, and near avoided crossings with higher lines the magnitude of $a$0 can increase strongly. In numerical simulations of flux-pulse-assisted readout, $a$1 changes from $a$2 MHz at $a$3 to $a$4 MHz near $a$5; in experiment, a more modest but still useful increase from $a$6 MHz to $a$7 MHz was demonstrated by pulsing to $a$8 (Stefanski et al., 2023, Stefanski et al., 2024).

A consequence is that fluxonium readout is not merely a problem of choosing $a$9, $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$0, and drive power. It is also a problem of navigating a flux-dependent network of virtual and real transitions. This suggests why flux pulsing, shelving, and explicit modeling of state transitions have become central parts of the field.

2. Resonator-coupled implementations

The canonical implementation is dispersive readout with an individual resonator per qubit. In a 2021 fluxonium processor, each qubit was coupled to its own quarter-wave coplanar-waveguide resonator with $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$1 GHz, $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$2 MHz, and an inferred $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$3 MHz; the noncomputational state $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$4 hybridized with one resonator photon strongly enough to help generate the observed dispersive shift. A tantalum-based high-coherence device instead used a 3D copper cavity with $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$5 GHz, $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$6 MHz, and measured $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$7 MHz at half flux bias. Granular-aluminum fluxonium has also been operated with a lumped-element LC resonator at $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$8 GHz and $$H\approx \hbar\omega_r a^\dagger a+\tfrac12\hbar\omega_q\sigma_z+\hbar\chi a^\dagger a\,\sigma_z.$$9 MHz, while scalable-architecture studies have proposed quarter-wave resonators in the $\chi$0 GHz range with $\chi$1 MHz and $\chi$2 MHz (Bao et al., 2021, Bothara et al., 28 Jan 2025, Gusenkova et al., 2020, Nguyen et al., 2022).

Implementation	Hardware	Reported or targeted readout parameters
Processor-scale fluxonium (Bao et al., 2021)	Quarter-wave CPW resonator per qubit	$\chi$3 GHz, $\chi$4 MHz, $\chi$5 MHz, contrast $\chi$6
Tantalum fluxonium (Bothara et al., 28 Jan 2025)	3D copper cavity, reflection readout	$\chi$7 GHz, $\chi$8 MHz, $\chi$9 MHz, $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$0 with JPA
Granular-Al fluxonium (Gusenkova et al., 2020)	Lumped-element LC resonator	$$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$1 GHz, $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$2 MHz, nearly flat transition rates up to $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$3
Scalable architecture study (Nguyen et al., 2022)	Four quarter-wave resonators on common bus	$$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$4 GHz resonators, $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$5 MHz, 4:1 multiplexing, no Purcell filter required

The scalable design literature treats the large detuning between a low-frequency fluxonium qubit and a $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$6 GHz resonator as a structural advantage. Because $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$7, Purcell decay of the low-frequency $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$8 transition is predicted to be negligible, and no dedicated Purcell filter is required. The same studies propose resonator spacing by $$\chi \simeq g_r^2\left(\frac{1}{\Delta_{ge}}-\frac{1}{\Delta_{ef}}\right),$$9, 4:1 shared-bus multiplexing, and wiring overhead of $\Delta_{ge}=\omega_q-\omega_r$0 line per 4 qubits, with high-SNR discrimination in $\Delta_{ge}=\omega_q-\omega_r$1 ns as a design target rather than an experimental benchmark (Nguyen et al., 2022).

3. Readout pulse sequences, shelving, and reset integration

The simplest protocol is single-tone homodyne measurement near the resonator frequency. In the 2021 processor, readout was performed with a single-tone homodyne measurement at $\Delta_{ge}=\omega_q-\omega_r$2; no detailed pulse envelope or microwave power was specified, but the authors noted that $\Delta_{ge}=\omega_q-\omega_r$3 ns “allows for fast readout.” The same platform integrated an active red-sideband reset: the $\Delta_{ge}=\omega_q-\omega_r$4 sideband transition was driven, the resonator then decayed with time constant $\Delta_{ge}=\omega_q-\omega_r$5 ns to $\Delta_{ge}=\omega_q-\omega_r$6, and a simultaneous short flux pulse moved the qubit slightly off sweet spot to lift the parity-forbidden selection rule. The measured post-reset ground-state population was $\Delta_{ge}=\omega_q-\omega_r$7 (Bao et al., 2021).

Readout is also tightly linked to feedback electronics. In an FPGA-based platform for granular-Al fluxonium, the readout pulse had rectangular envelope and duration $\Delta_{ge}=\omega_q-\omega_r$8 ns, the state classifier was a linear discriminant analysis threshold $\Delta_{ge}=\omega_q-\omega_r$9, and the measured platform latency from the last ADC sample to the first conditioned DAC sample was $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$0 ns. This enabled an active-reset sequence approximately $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$1s long with reset fidelity $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$2, reducing the excited-state population from $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$3 to $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$4 (Gebauer et al., 2019).

For low-frequency or weakly dispersive regimes, fluxonium readout often uses shelving into a more visible manifold. In heavy fluxonium with $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$5 MHz, plasmon-assisted readout drove $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$6 with a $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$7-pulse of length $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$8 ns and then discriminated $\Delta_{ef}=(\omega_e-\omega_f)-\omega_r$9 from $f_{01}/2\pi\approx0.56$0 through the resonator; the observed single-shot discrimination fidelity was $f_{01}/2\pi\approx0.56$1. In a MHz-frequency heavy fluxonium with $f_{01}/2\pi\approx0.56$2 MHz, the direct $f_{01}/2\pi\approx0.56$3 dispersive contrast was too small, so the protocol mapped $f_{01}/2\pi\approx0.56$4 with a 64 ns $f_{01}/2\pi\approx0.56$5-pulse and read out the $f_{01}/2\pi\approx0.56$6 manifold using a 600 ns pulse; corrected state-preparation fidelities were $f_{01}/2\pi\approx0.56$7 for both $f_{01}/2\pi\approx0.56$8 and $f_{01}/2\pi\approx0.56$9 (Zhang et al., 2020, Najera-Santos et al., 2023).

Historically, heavy-fluxonium work had already combined cavity-assisted readout and direct fluorescent readout. In that context, cavity-assisted dispersive readout reached $\approx0.4$0 in 500 ns, while direct fluorescent readout reached SNR $\approx0.4$1 in $\approx0.4$2s and fidelities $\approx0.4$3 (Earnest et al., 2017).

4. Flux-pulse-assisted and synchronized-flux readout

Flux pulsing has become the principal route to faster fluxonium readout because it exploits rather than suppresses the flux dependence of $\approx0.4$4. Theoretical work proposed moving the qubit from the sweet spot at $\approx0.4$5, where $\approx0.4$6 MHz, to a readout bias near $\approx0.4$7, where $\approx0.4$8 MHz. With a 50 ns linear ramp and $\approx0.4$9, the simulated SNR at $f_{01}(\Phi)$0 ns improved by $f_{01}(\Phi)$1 for $f_{01}(\Phi)$2 and by $f_{01}(\Phi)$3 for $f_{01}(\Phi)$4; the separation error $f_{01}(\Phi)$5 fell below $f_{01}(\Phi)$6 by $f_{01}(\Phi)$7 ns at $f_{01}(\Phi)$8 (Stefanski et al., 2023).

That proposal was subsequently realized experimentally without a quantum-limited parametric amplifier. During readout, a square flux pulse of amplitude $f_{01}(\Phi)$9 with 50 ns rise and fall edges shifted the device from a sweet-spot dispersive shift $a$00 MHz to a flux-pulsed value $a$01 MHz. With a readout tone at $a$02 GHz and $a$03 photons, the measured assignment fidelity reached $a$04 at $a$05 ns. From histogram fits, the SNR-limited fidelity corresponded to $a$06 at 360 ns, while the measured performance was limited chiefly by imperfect state initialization and relaxation during readout. The same work reported $a$07 and identified the flux-pulsed protocol as the fastest reported readout of a fluxonium qubit (Stefanski et al., 2024).

A more refined variant is synchronized-flux readout, designed to avoid state transitions localized in frequency space. In that approach the flux waveform is chosen so that

$a$08

with $a$09 following the cavity build-up and ring-down dynamics. Implemented with a 1-GHz-bandwidth AWG with $a$10 ns resolution and rise/fall shaping matched to $a$11 ns, this method avoided TLS crossings during the transient photon dynamics. After optimization over $a$12, $a$13, and drive frequency, and with post-selection of preparation, the net single-shot fidelities were $a$14 at $a$15s and $a$16 at $a$17s, compared with $a$18 and $a$19 at fixed bias. In the flux-compensated case, approximately $a$20 of the total error arose from $a$21 and $a$22 from $a$23 (Li et al., 19 Jul 2025).

5. Fidelity, QNDness, and high-photon-number operation

The best experimentally documented resonator-based fluxonium readouts now span a broad regime of speed and fidelity. In a tantalum-based device, single-shot assignment fidelity was $a$24 without a JPA and $a$25 with a JPA, using $a$26 and $a$27s without the JPA, or $a$28 and $a$29 ns with the JPA. The same system measured a QND repeatability fidelity

$a$30

At optimum power, the error budget was approximately $a$31 SNR-limited discrimination error without JPA and $a$32 with JPA, plus $a$33 state preparation and thermal population, $a$34 measurement-induced mixing, and $a$35 leakage (Bothara et al., 28 Jan 2025).

Other experiments show that fluxonium can remain near-QND at photon numbers far above those commonly used in transmon readout. In a granular-aluminum device, direct monitoring of quantum jumps found that both $a$36 and $a$37 remained flat within statistical error up to $a$38, with $a$39 kHz and $a$40 kHz. Although $a$41 decreased by $a$42 as $a$43 rose from 1 to $a$44, the SNR still grew monotonically, and the measurement time needed for target SNR $a$45 dropped from $a$46s at $a$47 to $a$48 ns at $a$49. In the same platform, feedback-assisted state preparation at $a$50 achieved $a$51 ground-state fidelity and $a$52 excited-state fidelity without a JPA; with a JPA, the excited-state fidelity rose to $a$53 (Gusenkova et al., 2020).

The contrast between platforms is notable. The 2021 processor reported an $a$54 readout contrast but did not publish a single-shot fidelity, explicit SNR, quantitative QND figure, or full readout error budget. By contrast, later work increasingly decomposed the infidelity into assignment, thermal, mixing, leakage, and transition components. This progression suggests that fluxonium readout matured from proof of distinguishability into a discipline organized around microscopic error accounting (Bao et al., 2021, Bothara et al., 28 Jan 2025).

6. Measurement-induced transitions, TLSs, and array-mode effects

A recurrent assumption is that increasing readout power mainly improves discrimination. Fluxonium experiments show a more device-dependent picture. In one granular-Al device, transition rates were essentially flat up to $a$55; in others, resonator photons induced substantial state evolution within and outside the computational manifold (Gusenkova et al., 2020, Bista et al., 29 Jan 2025, Zwanenburg et al., 16 Jun 2026).

The 2025 study of readout-induced leakage measured this explicitly. In Device A, $a$56 fell from $a$57 to $a$58 as $a$59 increased from 0 to 20, while $a$60 dropped much faster, reaching $a$61 already near $a$62, then briefly rising to $a$63 around $a$64 before falling again. The observed nonmonotonicity could not be explained by the bare fluxonium-resonator system alone; the best fit required a weakly coupled TLS with $a$65 MHz and $a$66 MHz (Bista et al., 29 Jan 2025).

A 2026 full-flux-range MIST study expanded this picture. It experimentally identified eleven distinct $a$67 regions with enhanced MIST. Six were explained by avoided crossings in a bare fluxonium-resonator description, while five additional regions required inclusion of the two lowest array modes of the superinductor. The same work reported excellent agreement between experiment and branch/Floquet analysis and concluded that array modes can dominate MIST at certain flux points (Zwanenburg et al., 16 Jun 2026). Closely related theory showed that these parasitic MIST processes can occur at relatively low readout drive powers, can leave finite occupation in an internal mode after the measurement, and can contribute to excess qubit dephasing even after the readout pulse is complete. The proposed mitigations were frequency allocation and coupling engineering: maximize detuning between array modes and the readout mode, and choose $a$68 to avoid low-order resonance conditions (Singh et al., 2024).

The synchronized-flux literature places TLS-induced transitions and MIST in a common framework. There the total non-QND error per shot is modeled as

$a$69

and TLS-induced errors appear as “hyperbolic fringes” in the $a$70 plane defined by

$a$71

This model captures why avoiding transient crossings during cavity ring-up and ring-down can be as important as the static readout point itself (Li et al., 19 Jul 2025).

7. Alternative modalities and broader applications

Not all fluxonium readout is dispersive. A 2025 experiment demonstrated non-demolition fluorescence readout and unconditional reset without employing a resonator. The device used a planar CPW stub filter with center frequency $a$72 GHz, 1 dB bandwidth $a$73 GHz, and more than 30 dB attenuation below 1 GHz, thereby suppressing decay on $a$74 MHz while enhancing the $a$75 readout transition at $a$76 GHz. The engineered decay rate was $a$77 MHz, the native $a$78 was $a$79s, the under-readout $a$80 was $a$81s, and the resulting QNDness metric $a$82 was $a$83. With 15 $a$84s integration and JPA amplification, the single-shot SNR was $a$85. The same platform implemented all-microwave unconditional reset with $a$86 fidelity in 200 ns and $a$87 in 250 ns (Watanabe et al., 22 Apr 2025).

Hybrid and application-specific variants extend the notion of fluxonium readout beyond direct cavity probing. In a fluxonium-transmon-fluxonium architecture, a 50 ns cross-resonance $a$88 pulse maps the fluxonium state onto a central transmon, after which standard transmon dispersive readout over $a$89 ns is used; the coherent mapping error is $a$90, and the resulting non-demolition assignment fidelity is $a$91 in the coherent limit, with transmon readout fidelity typically $a$92 (Dimitrov et al., 9 Sep 2025). In magnetic-field-compatible hybrid fluxonium, spectroscopy in fields up to $a$93 T was used to position the device as a readout circuit for topological qubits; the work did not report time-domain single-shot metrics, but it did demonstrate spectroscopic linewidths below $a$94 MHz at $a$95 T and explicitly framed the fluxonium as a persistent-current-based readout element for Majorana parity proposals (Pita-Vidal et al., 2019).

The most radical departures from cQED remain theoretical. Simulations of a single-fluxon readout architecture predict readout in less than 1 ns, without an input microwave tone, using state-dependent transmission or reflection of a ballistic fluxon at an interface containing the fluxonium qubit. In the mixed quantum-classical simulations, the reported backaction on the qubit was $a$96 (Wustmann et al., 26 Apr 2025). A plausible implication is that future classifications of fluxonium readout may be organized less by “dispersive versus nondispersive” than by whether the measurement channel is cavity-mediated, bath-engineered, ancilla-mediated, or ballistic.

Across these variants, the central problem remains consistent: to convert fluxonium’s multilevel, flux-tunable structure into a large and rapidly acquired measurement signal without activating unwanted transitions. The most successful solutions either exploit that structure directly, as in flux-pulsed and fluorescence readout, or offload the final discrimination to a more conventional subsystem, as in ancilla-mediated schemes.