Tunable Nonlinear Purcell Filters

• Frequency-tunable nonlinear Purcell filters are engineered environments that dynamically adjust the local density of states to suppress unwanted decay during qubit idle periods while enhancing measurement fidelity.
• They employ nonlinear mechanisms such as Kerr effects and saturable absorption to differentiate between weak noise fields and strong control signals, ensuring adaptive response under varying drive conditions.
• Implementations span superconducting circuits and nanophotonic platforms, offering scalable, multiplexed solutions that optimize trade-offs between speed, coherence, and radiative extraction.

Frequency-tunable nonlinear Purcell filters are tunable electromagnetic environments that reshape spontaneous emission, measurement backaction, and radiative extraction by controlling the spectral overlap between an emitter and its dissipative bath. In superconducting circuits, they are typically interposed between a qubit-associated resonator and a transmission line to suppress Purcell decay while preserving fast readout; in nanophotonic implementations, related structures modify the local density of states (LDOS) or modal density of states through Kerr, saturable, or plasmonic mechanisms to switch fluorescence enhancement and spontaneous-emission channels. Across these settings, the shared objective is dynamic control of the admittance or LDOS seen by the emitter, using flux bias, drive amplitude, optical pump intensity, or electrostatic gating as the tuning knob [(Whittaker et al., 2014); (Sunada et al., 2023); (Jahani et al., 2018); (Gruber et al., 2 Dec 2025)].

1. Physical principle

A Purcell filter is an engineered spectral environment that suppresses undesired emission at one frequency while retaining access to measurement or extraction channels at another. In circuit QED, this is usually expressed as control of the effective linewidth presented to a qubit through a readout chain. For a single linear mode of linewidth $\kappa$ coupled to a qubit with coupling $g$ and detuning $\Delta$, the qubit relaxation channel scales as

$$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$

or, more exactly,

$$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$

In nanophotonic formulations, the same phenomenon is written as LDOS engineering through

$$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$

The common structure is frequency selectivity in the environment seen by the emitter (Sunada et al., 2023, Jahani et al., 2018).

The additional qualifiers “frequency-tunable” and “nonlinear” identify two distinct control layers. Frequency tunability moves the relevant passband, stopband, notch, or cavity resonance in situ. Nonlinearity makes the filter response amplitude dependent, so that weak residual fields and strong measurement or control fields do not see the same transfer function. In practice, this can mean a flux-tunable Josephson cavity whose resonance and coupling vary together, a Kerr/Duffing resonator whose effective linewidth changes under drive, a saturable artificial atom that blocks spontaneous emission while becoming transparent to strong control pulses, or an ENZ metamaterial whose Kerr-shifted topological transition changes the LDOS abruptly [(Whittaker et al., 2014); (Iakoupov et al., 2022); (Sunada et al., 2023); (Jahani et al., 2018)].

2. Flux-tunable Josephson implementations in circuit QED

A canonical superconducting realization is the rf SQUID phase-qubit architecture of “Tunable-Cavity QED with Phase Qubits,” where an rf SQUID phase qubit is inductively coupled to a single-mode, flux-tunable lumped-element cavity. The coupled system is modeled by the Jaynes–Cummings Hamiltonian

$$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$

with bare qubit and cavity frequencies $\omega_{01}$ and $\omega_c$. Because the phase qubit is multilevel, the relevant dispersive shift is not the two-level result but the three-level expression

$$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$

where $g$0 and $g$1. In this device, cavity flux tuning changes both detuning and coupling: lowering $g$2 increases $g$3, while moving the cavity far from the qubit enlarges $g$4 and suppresses Purcell loss. The Purcell rate is modeled as

$g$5

Dynamic operation therefore places the cavity at high frequency during coherent evolution, where $g$6 is smaller and $g$7 is larger, and then shifts it to a lower frequency for readout, where $g$8 is larger. The reported cavity linewidth reaches $g$9 MHz, corresponding to a response time $\Delta$0 ns, and the maximum measured qubit lifetime reaches $\Delta$1s, attributed to dielectric loss with $\Delta$2 rather than cavity-mediated decay (Whittaker et al., 2014).

The same paper makes explicit that the cavity’s Josephson-junction nonlinearity is intrinsic to the filtering function. The cavity frequency depends on flux through a flux-tunable Josephson inductance $\Delta$3, while a series inductance $\Delta$4 reduces Josephson participation and flattens the frequency curve near its maximum. Power-dependent skewed Lorentzian resonances are observed, but the emphasis is not on strong bifurcation; rather, the cavity behaves as a reconfigurable Purcell-mitigation element whose nonlinearity and tunability jointly reshape the qubit’s radiative environment (Whittaker et al., 2014).

A later multi-qubit architecture pushes the same idea into shared readout and reset hardware. In “Flexible Readout and Unconditional Reset for Superconducting Multi-Qubit Processors with Tunable Purcell Filters,” the filter is a $\Delta$5 coplanar-waveguide resonator with a midpoint SQUID, flux tuned from $\Delta$6 to $\Delta$7 GHz. Its weak Kerr nonlinearity is measured as $\Delta$8 MHz, and its linewidth to the line is $\Delta$9 MHz. The filter is shared by three readout resonators, and the resonator’s effective external linewidth is modeled as

$$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$0

By moving the filter onto the readout band during measurement and away from it during idle, the architecture directly programs $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$1 while also reducing photon-noise dephasing and Purcell loss in idle periods. The same filter is also used as a fast dissipative bath for reset, enabling unconditional reset of both $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$2 and $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$3 within $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$4 ns with error rate $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$5, and $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$6-only reset in $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$7 ns (Xiao et al., 9 Jul 2025).

3. Drive-activated and saturable nonlinear filtering

A distinct line of work uses nonlinearity not merely to tune the center frequency, but to make the filter automatically respond differently to weak noise and strong readout or control fields. In “Photon-noise-tolerant dispersive readout of a superconducting qubit using a nonlinear Purcell filter,” the filter is a $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$8 resonator interrupted by a SQUID and galvanically connected to the readout line. The measured parameters are $$\Gamma_P \approx \left(\frac{g}{\Delta}\right)^2 \kappa,$$9 GHz, $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$0 GHz, $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$1 MHz, $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$2 MHz, $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$3 GHz, and filter anharmonicity $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$4 GHz. At low power and near resonance, the filter and readout resonator hybridize into two modes with equal linewidths $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$5, so the idle-state effective linewidth greatly exceeds $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$6 and suppresses dephasing from residual photons. Under strong readout drive, the filter’s Kerr shift detunes it from the readout resonator, reducing the effective coupling seen at the readout frequency. The resulting dephasing and measurement rates are

$$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$7

Experimentally, the noise tolerance is enhanced by a factor of $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$8 relative to a linear filter, and the measurement rate is enhanced by another factor of $$\Gamma_P = \frac{g^2 \kappa}{\Delta^2 + (\kappa/2)^2}.$$9 by exploiting bifurcation. Single-shot readout with a $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$0-ns pulse reaches $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$1 assignment fidelity and $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$2 QND fidelity (Sunada et al., 2023).

The saturable-filter approach of “Saturable Purcell filter for circuit quantum electrodynamics” uses a different nonlinear mechanism. A second artificial atom, a flux-tunable transmon acting as a Josephson quantum filter, is placed directly in the same transmission line used for both measurement and control. It is positioned at approximately half a wavelength at the qubit frequency, $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$3, to produce destructive interference and a spectral notch for emission at $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$4. The baseline Purcell decay without the filter is reported as $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$5 kHz, while the filter-induced dark-state fidelity is

$$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$6

For $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$7 MHz, the reported value is $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$8. Under strong control fields the filter saturates and effectively switches off, allowing resonant control through the same line. The paper reports average $$F_P(\omega)=\frac{\Gamma(\omega)}{\Gamma_0(\omega)}=\frac{\rho(r,\omega)}{\rho_0(\omega)}=\frac{6\pi c}{\omega}\,\mathrm{Im}\!\left[\mathbf{p}^* \cdot \mathbf{G}(r,r;\omega)\cdot \mathbf{p}\right].$$9-gate fidelities of $$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$0–$$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$1 for simple Gaussian-filtered rectangular pulses, $$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$2 after about $$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$3 optimal-control iterations, and $$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$4 after about $$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$5 iterations (Iakoupov et al., 2022).

These two circuit-QED strategies clarify that “nonlinear Purcell filter” is not a single device class. It may denote a Kerr/Duffing element whose bandpass self-adjusts under readout power, or a saturable absorber-like artificial atom whose stopband disappears under strong control. A plausible implication is that nonlinearity is valuable precisely when idle protection and driven accessibility must coexist on the same hardware path.

4. Optical and plasmonic realizations

Outside superconducting microwave circuits, frequency-tunable nonlinear Purcell filtering appears as active control of LDOS rather than qubit protection. “Switching Purcell effect with nonlinear epsilon-near-zero media” studies Ag/TiO$$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$6 hyperbolic metamaterial slabs near an ENZ frequency, where the optical isofrequency surface changes topology. The relevant nonlinear mechanism is Kerr-induced shifting of the effective permittivity,

$$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$7

which moves the ENZ condition and switches evanescent-wave transmission. In the Ag/TiO$$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$8 multilayer with silver filling fraction $$H = H_q + H_c + \hbar g\left(a^\dag \sigma^- + a \sigma^+\right),$$9, period $\omega_{01}$0 nm, and slab thickness $\omega_{01}$1 nm, a p-polarized pump at $\omega_{01}$2 nm, $\omega_{01}$3, and $\omega_{01}$4 GW/cm$\omega_{01}$5 suppresses the Purcell factor from $\omega_{01}$6 to $\omega_{01}$7, a $\omega_{01}$8 reduction. Away from ENZ, for example at $\omega_{01}$9 nm, the Purcell modulation is reported as $\omega_c$0. Finite-difference time-domain simulations show that pulse widths $\omega_c$1 fs reach the steady-state nonlinear response and that the switching speed is sub-picosecond (Jahani et al., 2018).

A more recent plasmonic implementation uses acoustic graphene plasmons. In “Tunable giant Purcell enhancement of quantum light emitters by means of acoustic graphene plasmons,” the resonator is a nanogap cavity defined by a graphene sheet and a $\omega_c$2-nm Ag nanocube, with an hBN/WS$\omega_c$3/hBN spacer of thickness $\omega_c$4 in the $\omega_c$5–$\omega_c$6 nm range. The AGP resonance depends on graphene Fermi energy $\omega_c$7, gap thickness $\omega_c$8, and number of graphene layers, giving real-time electrical tunability. The paper reports near-square-root scaling of AGP frequency with $\omega_c$9 and with $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$0. In the mid-infrared, the structure reaches $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$1 with quantum efficiency up to $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$2 for high-mobility graphene; at telecom wavelength $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$3m, it reports $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$4, quantum efficiency $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$5 for high-mobility graphene, and an on–off radiative-enhancement ratio of $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$6 dB when $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$7 is moved from $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$8 to $$\chi = \frac{(g^2/\Delta_{01})}{\left(1+\Delta_{01}/\alpha\right)},$$9 eV. For an erbium emitter inside single-layer WS$g$00, the E1 lifetime is reduced from $g$01s to $g$02 s, including quantum efficiency. The same platform is also analyzed for E2, E3, and two-photon spontaneous-emission channels (Gruber et al., 2 Dec 2025).

These optical and plasmonic examples use figures of merit different from those of superconducting readout hardware—Purcell factor, QE, and on–off radiative enhancement rather than $g$03, $g$04, or readout fidelity—but they implement the same operational idea: active spectral shaping of emission channels in a narrow, tunable band. This suggests a broad cross-platform definition of the topic in which a “Purcell filter” is not restricted to a microwave impedance transformer.

5. Optimization criteria, readout windows, and reset protocols

A central design problem is to choose the filter setting that maximizes information extraction while preserving coherence. In the 2014 tunable-cavity phase-qubit architecture, readout optimization is expressed through the condition $g$05, with

$g$06

while the drive photon number must satisfy the critical-photon bound

$g$07

The same work emphasizes dynamic timing: the cavity stays at a “safe” high frequency during coherent evolution and is shifted only during the brief readout window, taking advantage of the measured $g$08 ns cavity response time (Whittaker et al., 2014).

In the 2023 nonlinear-filter readout experiment, the operative quantity is the drive-dependent $g$09 of the readout chain. Under that paper’s conventions, $g$10 is maximized when $g$11, whereas the idle state deliberately uses $g$12 to suppress dephasing per stray photon. The experiment also exploits bifurcation of the nonlinear filter so that the two qubit states occupy different response branches. After measurement, the hybridized modes relax rapidly: the resonator empties from $g$13 to $g$14 in about $g$15 ns, so no active resonator reset is needed (Sunada et al., 2023).

In the 2025 shared-filter architecture, the non-steady-state homodyne analysis gives the standard optimum $g$16 for the chosen convention, with cavity ring-up time $g$17. The paper demonstrates $g$18–$g$19 readout with $g$20 ns integration and SNR $g$21, $g$22–$g$23 readout with $g$24 ns integration and SNR $g$25, and $g$26–$g$27 readout with $g$28 ns integration, SNR $g$29, and fidelity $g$30, all without JPA or TWPA and with a small dispersive shift $g$31 MHz. The same hardware supports reset by a qubit–coupler swap followed by a coupler–filter swap while the filter is parked at $g$32 GHz and the readout resonators remain near $g$33–$g$34 GHz. Reported durations are about $g$35 ns for the adiabatic qubit–coupler swap and about $g$36 ns for the adiabatic coupler–filter swap, yielding single-cycle unconditional reset of $g$37 and $g$38 in $g$39 ns and repeated-cycle error below $g$40 within $g$41 ns (Xiao et al., 9 Jul 2025).

The differing optima—$g$42 in one convention, $g$43 in another, and again $g$44 in a non-steady-state homodyne treatment—do not indicate a disagreement in physical objective. Rather, they reflect different definitions of $g$45 and different response models. In each case, the filter is adjusted so that the measurement channel is strong only when measurement is intended.

6. Fluctuations, limitations, and scalability

Tunable and nonlinear filters introduce a second design problem: the filter itself can fluctuate. “Frequency Fluctuations in Tunable and Nonlinear Microwave Cavities” models the measured scattering response as an average over resonance-frequency jitter,

$g$46

with effective fluctuation scale

$g$47

If these fluctuations are not included in the fit model, damping rates can appear to depend spuriously on the tuning parameter. The paper gives practical thresholds for keeping apparent linewidth and coupling biases below $g$48: $g$49 for Gaussian fluctuations, $g$50 at tuning sweet spots where quadratic fluctuations dominate, and $g$51 for the quantum-limited Kerr case at very low photon number. For $g$52, maintaining both biases below $g$53 requires $g$54 (Brock et al., 2019).

These fluctuation results sharpen several trade-offs already present in the platform-specific studies. In circuit QED, increasing $g$55 or moving the filter into stronger resonance improves bandwidth and readout speed but can increase Purcell loss or photon-noise sensitivity if the device is not detuned during idle. In saturable filters, larger $g$56 broadens the protective notch and accelerates bright-state decay, but it also raises the required control power for saturation. In ENZ metamaterials, metallic absorption and the imaginary part of $g$57 limit modulation depth and remove bistability when two-photon absorption is included. In AGP cavities, graphene absorption, Ag ohmic loss, nanogap tolerances, and emitter placement determine whether the bright radiative mode is accessed efficiently (Sunada et al., 2023, Iakoupov et al., 2022, Jahani et al., 2018, Gruber et al., 2 Dec 2025).

A recurrent misconception is that a Purcell filter is necessarily fixed and linear. The cited literature shows instead a spectrum of implementations: weakly nonlinear flux-tunable Josephson cavities used primarily for reconfigurable Purcell mitigation, Kerr/Duffing filters whose effective linewidth self-adjusts under readout drive, saturable artificial-atom filters that suppress idle decay while passing strong control pulses, ENZ slabs whose LDOS changes under femtosecond optical pumping, and AGP resonators whose enhancement band moves under electrostatic gating [(Whittaker et al., 2014); (Sunada et al., 2023); (Iakoupov et al., 2022); (Jahani et al., 2018); (Gruber et al., 2 Dec 2025)].

Scalability is a major reason these devices are studied. The tunable cavity of the 2014 phase-qubit work was already proposed as a way to reduce residual bus coupling and cavity-induced dephasing in multi-qubit systems. The saturable-filter proposal explicitly states that combining the filter with frequency multiplexing can enable control and measurement of several qubits using a single Purcell-filtered transmission line. The 2025 shared-filter architecture implements one filter for three readout resonators, with eight filters serving twenty-four qubits on a flip-chip processor. The 2023 nonlinear-filter readout study also states compatibility with scalable, multiplexed readout [(Whittaker et al., 2014); (Iakoupov et al., 2022); (Xiao et al., 9 Jul 2025); (Sunada et al., 2023)].

Taken together, these developments define frequency-tunable nonlinear Purcell filters as a general strategy for programmable radiative engineering. Their technical forms differ—bandpass admittance shaping, destructive interference and dark-state formation, Kerr-induced transfer-function reconfiguration, ENZ topological switching, or gate-tuned plasmonic mode selection—but each uses tunability plus nonlinearity to separate idle protection from driven functionality.