Fast Navigation in a Large Hilbert Space Using Quantum Optimal Control

Arthur Larrouy, Sabrina Patsch , Rémi Richaud, Jean-Michel Raimond , Michel Brune , Christiane P. Koch , and Sébastien Gleyzes 1,* Laboratoire Kastler Brossel, Collège de France, CNRS, ENS-Université PSL, Sorbonne Université, 11, place Marcelin Berthelot, 75005 Paris, France Theoretische Physik, Universität Kassel, Heinrich-Plett-Straße 40, 34132 Kassel, Germany Dahlem Center for Complex Quantum Systems and Fachbereich Physik, Freie Universität Berlin, Arnimallee 14, 14195 Berlin, Germany


I. INTRODUCTION
Quantum information science exploits the essential quantum features of light and matter for the design of devices with applications in computing, sensing, or communication [1]. The building blocks of any quantum device are qubits as information carriers, and their operation requires the ability to prepare, manipulate, and read out their state [2]. The simplicity of the basic concept is, however, in contrast with the complexity encountered when implementing it in an actual physical platform. Both in atom-based and solid-state devices it becomes more and more challenging to precisely control the quantum state as the size of the system increases. For example, higher-order terms of the Hamiltonian introduce nonlinearities that limit the fidelity of basic operations [3,4]. The complexity of the system spectrum increases, making it difficult to address individually quantum states in a short time [5,6]. At the same time, decoherence occurs more rapidly, which requires to operate faster [7].
Quantum optimal control [8,9] is a versatile approach to address the challenge of fast navigation in a large Hilbert space. It allows one to design the time-dependent shape of experimental control knobs, electromagnetic fields, for instance, that accomplish a given task in a quantum system in the best possible way. The starting point is a performance index, such as the target state preparation fidelity, which is treated as a functional of the (yet unknown) shape of the electromagnetic control field. The essence of quantum optimal control theory is to maximize the performance index while minimizing the use of resources such as time, bandwidth, or power [8].
A large number of theoretical quantum control protocols cover most physical platforms considered for quantum technologies, notably neutral atoms [10], ions [11], photons [12,13], color centers in diamond [14,15], and superconducting qubits [16,17]. Quantum control experiments have progressed at a much slower pace and with a focus on one or two qubits or few-level systems [18][19][20][21][22]. For systems with larger Hilbert spaces, such as the quantum harmonic oscillator [23,24] or an assembly of trapped atoms [7,25], optimal control approaches start to replace conventional quantum engineering designs, providing a faster alternative [26] to existing preparation protocols [27]. Here, we go further and demonstrate how quantum optimal control can be used to navigate a Hilbert space of large dimension, finding a strategy to reach an arbitrary quantum state, for which no intuitive preparation method is known.
Our example is the manipulation of Rydberg atoms, which represent an attractive platform for quantum technologies [28]. In particular, Rydberg atoms in long-lived circular states, where the angular momentum projection takes its maximal value, are ideal for fundamental experiments on matter-field coupling [29]. They have also been suggested for use as logical states in a quantum computer [30,31] or as an optical-to-microwave interface [32,33]. Superpositions of circular states with opposite magnetic momentum or of states with very different electric dipoles are instrumental for quantum-enabled electrometers and magnetometers [4,34]. Use of circular states and their superpositions in quantum computing and sensing in particular requires the capability of fast and accurate state preparation.
In this paper, we make use of optimal control to coherently manipulate Rydberg atoms and prepare an arbitrary superposition of states in a complex Hilbert space inside the Stark manifold of rubidium. We first demonstrate a significant improvement of the preparation of the long-lived circular state in terms of fidelity and duration. Furthermore, we prepare mesoscopic superpositions of angular momentum eigenstates in the Rydberg manifold that cannot be created with standard protocols. This achievement is made possible by using quantum optimal control to tune the time-dependent phase and amplitude of the rf drive [35] that coherently manipulates the atoms inside the Rydberg manifold. Our work opens the way to generate arbitrary nonclassical superposition states in large Hilbert spaces.

II. RYDBERG ATOMS
We consider rubidium atoms in the presence of an electric field F. The Hilbert space to be navigated is sketched in Fig. 1, showing the rubidium levels with a principal quantum number n ¼ 52 and a positive magnetic quantum number m. The levels are arranged in a triangular shape with the circular state 52c at its tip [36]. For m > 2, the levels are almost hydrogenic, and the transitions between nearby levels are, to first order in F, degenerate at the Stark frequency ω n =2π ¼ 3nea 0 F=2h (250 MHz for . For m ≤ 2, the energy levels are shifted due to the electron penetration in the ionic core and the transitions can differ from a few megahertz to hundreds of megahertz with respect to ω n [37]. Transitions between different m sublevels couple to different polarizations of the electromagnetic field. An atom initially prepared in a state of the lowest diagonal of Fig. 1 will remain in this subspace (represented by thick lines on Fig. 1) when driven by a σ þ -polarized radio frequency at frequency ω rf ¼ ω n (blue arrows in Fig. 1). This subspace can be described as a large spin J ¼ 51=2, evolving on a generalized Bloch sphere, with the circular state at the north pole. For the hydrogenic levels (m > 2), the rf field induces a classical rotation of the spin on this sphere [33].
The experiment takes place in a plane-parallel capacitor providing the directing field F aligned with the Oz quantization axis (Fig. 2). The atoms are prepared at t¼0 in the lowest m ¼ 2 level (label m 2 on Fig. 1) by a combination of laser pulses [33] (see Appendix B). They are then manipulated by a rf field created by four electrodes circling the interval between the capacitor plates. We apply on them rf drives at ω rf =2π ¼ 250 MHz generating near the center of the structure a σ þ -polarized field. After having FIG. 1. Stark levels. Scheme of the energy levels for F ∼ 2.5 V=cm, sorted by m values for m ≥ 0. The circular state is labeled c. The levels of the lower diagonal are labeled m j for low-m states and e j for high-m states. Finally, e 0 1 is the Stark level with m ¼ 50 above the circular state. The laser excitation prepares the m 2 state. The atom is then driven by a σ þ rf field at frequency ω rf ¼ ω n (blue arrows). Since a σ þ field only couples levels with Δm ¼ þ1, the dynamics of the atom remains in the subspace of the manifold lower diagonal (levels represented as thick lines).
FIG. 2. Scheme of the experiment. A rubidium thermal beam (dark blue arrow), effusing from an oven, crosses a structure made of two plane-parallel electrodes A and B (blue) creating the vertical directing electric field F. They are surrounded by four ring electrodes (yellow, two of them are not represented). At the center O of the structure, the atoms interact with three laser beams (780 and 776 nm, σ þ polarized, red arrow; 1258 nm, π polarized, green arrow) that cross at 90°in the horizontal plane. At the end of the sequence, the atoms are detected in the state-selective fieldionization detector D.
interacted with the rf field, the atoms drift toward another capacitor, where they are detected by state-selective field ionization. The detector resolves levels of different manifolds or with very different m values in the same manifold. We measure the population of levels with overlapping fieldionization signals by applying additional probe microwave (MW) pulses, as described in Appendix C.

III. FAST CIRCULARIZATION
The standard procedure for rf-induced circular state preparation is a rapid adiabatic passage sequence [33], in which the electric field is slowly ramped down across F 0 in the presence of the rf field. Since the transition frequency ω 2;3 between 52m 2 and 52m 3 only differs from ω n by a few megahertz for F ≈ F 0 , scanning the field over 0.24 V=cm transfers the atoms from 52m 2 into 52c with a 99.5% efficiency, cf. Appendix D. To reach such a high efficiency, the adiabatic passage method has to be slow, with a preparation time in the few microsecond range at least. Such a long time can be detrimental, in particular for quantum sensing experiments [4,34].
The 52c state can be prepared much faster from the same 52m 2 level by setting F ¼ F 0 and applying a constant amplitude rf pulse [33]. The duration of the preparation (96 ns) is only limited by the applied rf amplitude. Figure 3 shows the time evolution of the populations of relevant states. The maximum transfer efficiency is limited to 77.4 (4)%, significantly lower than that of the adiabatic passage. Some of the population remains trapped in noncircular levels, due to the frequency shifts of the m ¼ 0 and m ¼ 1 levels induced by the quantum defects.
Optimal control allows us to combine fast preparation with high fidelity. We use as experimental knobs two timedependent voltages, V r ðtÞ and V i ðtÞ, generated by two outputs of an arbitrary waveform generator (Appendix E). They control rf mixers which modulate the amplitudes F r ðtÞ and F i ðtÞ of the two quadratures of the σ þ -polarized rf field: The process we seek to optimize is a state-to-state transfer from the initial state, 52m 2 , to the target state jΨ tgt i, 52c [35]. The success of the transfer is quantified by the realvalued target functional, where jΨðTÞi is the final state at the end of the protocol. A perfect match of final and target state, up to a global phase, is obtained if and only if J T takes its minimal value, J T ¼ 0. The final state depends implicitly, via its time evolution, on the external controls F r ðtÞ and F i ðtÞ. An optimal choice of F r=i ðtÞ can be calculated by minimization of the target functional [8]. Here, we use Krotov's method to minimize J T . It is a gradient-based optimization technique that ensures monotonic convergence of J T toward its minimum [38]. A detailed, hands-on introduction of the method is found in Ref. [39]. As with any gradient-based technique, the condition for J T to be extremal, ∇J ¼ 0, results in an update equation for the external controls, and evaluation of the gradient implies forward propagation in time of the initial state and backward propagation of the target state [8]. In practice, the functional is thus minimized iteratively by updating F r ðtÞ and F i ðtÞ until we reach a transfer efficiency of 99% in the simulation.
In our system, the rf pulse mainly couples m levels on the lowest diagonal in Fig. 1, but off-resonant excitation may also populate the levels on the second lowest diagonal. These levels are thus included in the calculation. To ensure that the optimized field is experimentally feasible, we need to include constraints on the spectral bandwidth and the maximum amplitude of the pulse. The simplest possible implementation-truncation of the optimized pulse to the allowed range of amplitude and frequency after each iteration step-has turned out to be sufficient to this end [35].
The optimized pulse is presented Fig. 4(a). The dynamics induced by the shaped pulse is discussed in detail in Ref. [35]. Briefly, the optimized pulse can be decomposed into two phases [separated by a vertical line in Fig. 4(c)]. The first 40 ns prepare a superposition of hydrogenic levels (m > 2) that corresponds to a spin coherent state (SCS) on the Bloch sphere, while leaving no spurious population in m ¼ 1 [Figs. 4(c) and 4(d)]. The last 73 ns rotate the SCS onto the circular state.
The optimization algorithm produces a pulse, in which the amplitude of the two quadratures is modulated even after τ > 40 ns. This corresponds to a slow oscillation of the direction of the rotation axis of J on the Bloch sphere. This modulation is not necessary for reaching the circular state. It results from the constraint on the spectral bandwidth that is imposed during the optimization. The physical insight in the dynamics under the optimized pulse allows us to remove the unnecessary modulation. We have checked numerically that, after t > 40 ns, these oscillations can be flattened out into a pulse of constant amplitude and wellchosen phase without losing the transfer efficiency into the circular state [Figs. 4(b) and 4(c)].
Using the rf mixer calibration (Appendix E), we calculate, from the flattened theoretical pulse, the voltage V r ðtÞ and V i ðtÞ to be applied in the experiment. We finally perform a "closed-loop" optimization, directly on the experimental signal. For the first 40 ns the two quadratures are rescaled by the same 0.95 factor, close to 1, in order to compensate for the rf amplitude calibration uncertainty. For the flat part of the pulse (τ > 40 ns), we finely tune the SCS rotation by independently adjusting the constant values of V r ðtÞ and V i ðtÞ in order to optimize the final circular state population (Appendix F). Figure 5 presents the experimental evolution of the Stark levels populations (dots) for the optimized pulse (abruptly interrupted after a variable duration τ). The transfer probability into the circular state now reaches 96.2(3)%, for a preparation time (113 ns) comparable to the duration of the square pulse (Fig. 3). The fact that only about 25 oscillations of the carrier rf wave perform such an efficient transfer is a remarkable result. The evolutions of the probed level populations are in good agreement with the numerical predictions (solid lines). Quantum control provides thus a considerable improvement over the usual circular state preparation methods.
We now test the coherence of the prepared state by checking that the process does not spoil a coherent superposition with a reference level. For practical reasons, we The discrepancy between the data and the simulation at the beginning of the optimized pulse is probably due to the finite bandwidth of the rf circuit. When we set V r and V i to zero at time τ, the rf has a finite ring-down time, which affects the measured population. The effect of this ring down is not visible at the end of the pulse, as it corresponds there to a small additional rotation of the spin, which is automatically compensated by the optimization procedure. Finally, we find P c ¼ 96.2ð3Þ%, P e 1 ¼ 0.20ð1Þ%, P e 2 ¼ 0.74ð4Þ%. We also measure P e 3 ¼ 0.15ð7Þ%, P e 4 ¼ 0.13ð8Þ%, P m 3 ¼ 0.08ð3Þ%, start from a superposition of 52c (reached by an initial rapid adiabatic passage) and the reference level 50c. This superposition is prepared by a π=2 mw pulse driving the 52c → 50c two-photon transition. A sequence of a time-reversed optimized rf pulse (Appendix F) and of a direct one, separated by a 10 ns delay, drives 52c into 52m 2 and back. A final π=2 mw pulse mixes again 52c and 50c and closes a Ramsey interferometer. The optimized time-reversed and direct pulses are shown in Fig. 6(a). Since the Stark frequency ω 50 =2π differs by only 10 MHz from ω rf =2π, the rf driving produces a spurious rotation of the spin associated to the n ¼ 50 manifold. By using two amplitude levels for the first 73 ns of the reversed pulse and the last 73 ns of the direct pulse, we ensure that each pulse returns 50c exactly onto itself, while preserving the optimal transfer between 52c and 52m 2 (Appendix F). Figure 6(b) presents the probability for finally detecting the atoms in 52c as a function of the relative phase of the two MW pulses. The fringe visibility is 0.891(6), a large value, limited by electric field noise. The atoms are transiently cast in a superposition of levels (52m 2 and 50c) with very different dipoles which is utterly sensitive to the electric field [4]. An extrapolation to zero electric noise leads to a visibility equal to 1 within the error bars (Appendix G). We thus conclude that the rf optimized pulse preserves coherences extremely well.

IV. CAT STATE PREPARATION
Optimal control opens much wider possibilities than merely improving circular state preparation. We use the same methodology as described above, simply replacing the target state jΨ tgt i by an equal weight superposition of the lowest level in the m ¼ 1 manifold (52m 1 ) with the circular state (52c). This superposition is a Schrödingercat-like state useful for quantum-enabled electrometry [4]. It is particularly challenging to prepare, as its preparation is equivalent to a π=2 pulse on a 50-photon transition.
We again design the rf pulse driving the desired state transfer using Krotov's optimization method. Figure 7  reaches the circular state, the rf is slowly switched off so that the part of the wave function in the dark state adiabatically transfers into a pure 52m 1 level.
Similarly to the case of circular state preparation, the qualitative understanding of the dynamics allows us to simplify the pulse. We replace the pulse in the second half of the protocol (τ > 80 ns) by a linearly increasing and decreasing amplitude of the control voltage of each quadrature (Appendix F). This enables us to tune the rotation angle of the spin coherent state (by independently changing the pulse area for each quadrature), while ensuring that the phase and amplitude of the rf field vary slowly enough for the other half of the wave function to adiabatically remain shelved in the instantaneous dressed state. Figure 8(a) shows the programmed pulse. Its shape includes minor compensations for the finite electronic bandwidth, and the parameters of the two linear amplitude ramps are adjusted to optimize the rotation of the SCS (Appendix F). Figure 8(b) shows the experimental evolution of the relevant level populations (dots). The preparation of the dressed state with a large component of 52m 1 is evidenced by the nearly constant population of 52m 1 after 60 ns. The final population balance between 52m 1 and 52c is excellent. The experiment is again in good agreement with the numerical model (solid lines).
In order to test the coherence of the cat state, we apply, after an adjustable delay Δt, a time-reversed preparation pulse. During the delay, the superposition state accumulates a phase in the frame rotating with the rf carrier frequency. The state at the end of the time-reversed pulse thus oscillates between 52m 2 and a state orthogonal to it as a function of Δt. Figure 8(c) shows the oscillations, with Δt, of the probability to find the atoms finally in 52m 2 . The visibility of the interference is high [0.80 (1)]. An extrapolation to zero electric field noise (Appendix G) leads to a visibility of 0.97 (2). We estimate that we prepare the expected state superposition with a remarkable 93% fidelity.

V. CONCLUSIONS
We have used quantum optimal control to prepare nontrivial Rydberg states with high fidelity in short times. The optimized pulses are very robust to the limited calibration of the control electronics. At the same time, they are comprehensible, which eases adaptation of the pulses to the experimental constraints. For example, the strategy for the cat state pulse consists in preparing a superposition of a spin coherent state and an eigenstate that is kept dark during the rotation of the spin coherent state. This understanding allows us to experimentally fine-tune the final rotation of the spin coherent state without affecting the final population of 52m 1 .
The preparation could be made even faster if more rf power is available in the experiment. The method could be extended to a larger variety of Rydberg quantum states, in particular by applying simultaneously σ þ -and σ − -polarized shaped pulses, with interesting applications for quantum-enabled electrometry and magnetometry [4,34]. More generally, our results attest for the power and reliability of quantum optimal control in a Hilbert space of a large dimension, with a complex combination of harmonic and anharmonic ladders. They open the way for quantum optimal control based state engineering in complex systems for applications in quantum science and technology.

ACKNOWLEDGMENTS
Financial support from the Agence Nationale de la Recherche under the project "SNOCAR" (167754) and the Studienstiftung des deutschen Volkes is gratefully acknowledged. This publication has received funding from the European Union's Horizon 2020 research and innovation programme under Grant Agreement No. 817482 (PASQuanS).

APPENDIX A: NUMERICAL OPTIMIZATION
In this appendix, we describe additional details of the pulse optimization (see also Ref. [35] for a comprehensive description).
In order to calculate the dynamics of the atoms, we first numerically compute all the eigenvalues of the Stark Hamiltonian (in the absence of rf field) for each magnetic quantum number m ≥ 0. For this diagonalization we consider all states with a principal quantum number of n ¼ 52 AE Δn. We use the known rubidium quantum defects up to l ¼ 7 [40] for describing the energy levels in zero electric field. We then compute the transition matrix elements between neighboring m and m þ 1 Stark levels.
We check the validity of the Hilbert space truncation by comparing the eigenstates and transition matrix elements for one value of Δn with an extended Hilbert space with Δn þ 1. For the chosen value of Δn ¼ 4, this method leads to an accuracy of the order of 10 −6 in the energy levels and transition matrix elements.
At the field strength F ¼ F 0 ¼ 2.5 V=cm, we find transition frequencies ω m;mþ1 between m and m þ 1: For m ≥ 3, all the transition frequencies ω m;mþ1 =2π are less than 1 MHz away from ω rf =2π (the difference being due to residual quantum defect shifts and the second-order Stark effect).

Fast circularization
The optimization of the desired state-to-state transfer in the Stark manifold of the rubidium Rydberg atom was performed using Krotov's method [35]. We have assumed the initial state to be the lower state of the 52 m ¼ 2 ladder of the Stark manifold. For this optimization, it is sufficient to take the lowest two diagonal ladders of the n ¼ 52 manifold into account and to operate thus in a Hilbert space of dimension 103 [35]. As a final check, we perform a numerical integration with the full, untruncated Hamiltonian using the optimized pulse. We find the same final population of the target state as in the truncated space within a precision of 5 × 10 −5 .
In the optimization, we take into account two experimental limitations. At each iteration of the optimization, we truncate the pulse bandwidth by a square window (frequency bounds 140 ≤ ω rf =2π ≤ 360 MHz, with sineshaped edges of 20 MHz width). This avoids getting variations of the quadrature amplitudes that would be too fast to be implemented experimentally. We also limit the pulse amplitude by truncating the pulse to the maximal experimentally available value F rf ¼ 5 V=m.
We then carry out the optimization for a given total pulse duration, targeting a preparation fidelity larger than 99%. The guess pulse has a flattop shape with sine-shaped edges with a rise and fall time of 10 ns. For too short pulse durations, the algorithm does not converge within the imposed limitations; it first reaches convergence for a duration of 113 ns.

Cat state preparation
The target state for this optimization is chosen to be ðj52m 1 i − j52ciÞ= ffiffi ffi 2 p . The maximal rf amplitude was kept at F rf ¼ 5 V=m, but the bandwidth of the pulse was slightly changed to 130 ≤ ω rf =2π ≤ 320 MHz with sineshaped edges of 50 MHz width, such that the center of the frequency window is a bit lower than for the circularization pulse. The guess pulse was similar to the circularization one. The duration of the pulse had to be increased to 150 ns for a successful optimization.

APPENDIX B: TIMING OF THE EXPERIMENT
The setup is depicted Fig. 2. A thermal beam of Rb, effusing from an oven, crosses a structure made of two planeparallel electrodes surrounded by four electrodes that form a ring around them. At the center of the structure, the atoms interact with three laser beams at 780, 776, and 1258 nm, resonant with the 5S 1=2 → 5P 3=2 , 5P 3=2 → 5D 5=2 , and 5D 5=2 → 52F transitions, respectively. The 780 and 776 nm propagate along the same direction, and are always on. The 1258 nm laser is sent perpendicular to the other laser beams, and is switched on for 1 μs at the beginning of the sequence. The beginning of the laser pulse sets the time origin t ¼ 0.
During the laser excitation, the quantization axis is defined by a small horizontal electric field along the axis of the 780 and 776 nm lasers, created by a voltage applied across two of the ring electrodes. This field partially lifts the degeneracy of the 52F level, and we set the 1258 nm laser to be on resonance with the transition toward the 52F sublevel with jmj ¼ 2. We also choose the polarization of the 780 and 776 nm lasers to be σ þ polarized and the 1258 nm laser to be π polarized with respect to the quantization axis. We thus prepare the atoms selectively in the state j52F; m ¼ 2i.
Once the atoms have been excited to the Rydberg state, the electric field is adiabatically rotated and ramped up to F ¼ 2.63 V=cm between t ¼ 2 μs and t ¼ 3 μs. It is finally aligned with Oz, which is the quantization axis for the rest of the experiment. It remains constant for 1.3 μs, and decreases down to F ¼ 2.39 V=m in 1.5 μs. It is finally kept constant for 1 μs. This variation of the electric field makes it possible to prepare the 52c circular state with a high purity by adiabatic passage if needed as initial state for the rest of the sequence. In this case, we ramp up at t ¼ 2.91 μs the amplitude of the rf field to a value of 4.8 V=m in 2.05 μs. We leave the amplitude constant for 0.1 μs, and ramp it down to zero in 2.05 μs. When decreasing F ¼ 2.63 V=cm down to F ¼ 2.39 V=cm, the Stark frequency in n ¼ 52 varies between 262 and 238 MHz and crosses the rf source frequency while the rf amplitude is near its maximum. Ramping down the electric field in the presence of the rf adiabatically transfers the atoms from 52m 2 into the circular state. If we choose not to apply the rf, the atoms remain in 52m 2 . At t ¼ 7 μs, we set the amplitude of the electric field to the value F 0 , and we start the optimized rf pulse at t ¼ 8.71 μs.

APPENDIX C: MEASUREMENT OF LEVEL POPULATIONS
The field-ionization detection method does not resolve Stark sublevels of the same manifold that have similar m values. Therefore, to measure the population of a given sublevel of the n ¼ 52 manifold after the rf pulse, we selectively transfer its population into a level inside the 51 or 50 manifolds with a MW π pulse, called "probe pulse," and count the number of atoms in the target manifold. However, different atomic levels may have different detection efficiencies (Fig. 10). As a result, we use specific normalization methods for each relevant level.

Low-m states
We measure the population of the levels 52m i (i ¼ 1, 2, 3) by applying, after the end of the rf pulse, a MW probe π pulse tuned on one of the 52m i → 51m 2 transitions. In order to avoid a strong count rate reduction due to the finite lifetime of 51m 2 , we then transfer the population of 51m 2 to 51c by adiabatic passage and count the number of atoms in the 51c state. This number is proportional to the population of the probed levels and to the efficiencies of the 52m i → 51m 2 π pulses. They are carefully calibrated and included in the calculation of populations in 52m i . These populations are obtained by using as a normalization factor the number of atoms counted in 52m 2 (via 51c) in the absence of rf pulse. Note that this method discriminates the states with m ¼ 2 and with m ¼ −2. Atoms initially prepared in the 52; m ¼ −2 state due to small laser polarization imperfection are transferred in 51; m ¼ −2 by the probe MW π pulse but are not transferred into 51c by the σ þ rf circularization pulse.
The 52m i → 51m 2 (i ¼ 1, 3) transitions have a large linear Stark effect. A slow drift of the static electric field (∼10 −4 V=m over an hour timescale) results in significant errors in the calibration of MW π pulse efficiencies (shaded area on Fig. 3 of the main text). For the cat state preparation (Fig. 8 of the main text), we normalize the population in 52m 1 by the number of atoms directly prepared in 52m 1 from 52m 2 using a resonant rf π pulse (efficiency 99% as seen on Fig. 9). This method is insensitive to the 52m 1 → 51m 2 MW probe pulse efficiency calibration.

High-m states
We measure the population of the 52c, 52e i (i ¼ 1, 2, 3, 4), and 52e 0 1 levels, by applying a probe microwave π pulse that selectively transfers one of these levels into the corresponding one in the 50 manifold [33], which is finally detected. For each of these levels, we calibrate a correction factor that takes into account the efficiency of the microwave π pulses and the relative detection efficiency of relevant levels in order to estimate their populations [41]. We normalize populations by the number of atoms that we detect with the 52c probe when we prepare the circular state by adiabatic passage. This method thus directly compares the efficiency of the preparation of the circular state using the optimized pulse with that of the adiabatic passage (estimated at ∼99.5%; see below), independently from the transfer efficiency of the circular state MW probe. Figure 10 presents the ionization signals of the atomic state with and without performing the adiabatic passage at t ¼ 2.91 μs. We clearly see that most of the population in the initial state has been transferred to high-m states, which ionize in larger fields. The residual counts at the position of the 52m 2 ionization peak (about 0.5%) most likely MHz rf field that is resonant with the 52m 2 − 52m 1 transition. The points represent the population P m 2 remaining in the 52m 2 state as a function of duration τ of the pulse. For τ ¼ 88 ns, we measure P m 2 ≈ 0.3%. The amplitude of the rf is chosen low enough so that the atoms cannot be transferred into 52m 3 (numerical simulations estimate that P m 3 < 0.5%). The pulse thus prepares 51m 1 with more than 99% efficiency.

APPENDIX D: ADIABATIC PASSAGE EFFICIENCY
correspond to atoms that have been prepared in the m ¼ −2 state (due to the excitation laser polarization imperfections) and do not interact with the σ þ radio frequency. However, our detection method using probes distinguishes between m ¼ AE2 (see above). The m ¼ −2 atoms thus do not contribute to the measured populations in the m ¼ 2 state. There are two possible effects that limit the efficiency of the adiabatic passage [33]. On the one hand, if the adiabaticity criterium is not fulfilled, some population can remain in the second last level of the energy ladder, 52e 1 . On the other hand, if the polarization of the radio frequency is not purely σ þ , some population can end up in the elliptical state 52e 0 1 , defined on Fig. 1. In order to precisely quantify the purity of the adiabatic circular state preparation, we have measured the population of the 52e 1 and 52e 0 1 states after the adiabatic passage using probes [33]. We find P e 1 ¼ 0.15ð1Þ% and P e 0 1 ¼ 0.39ð3Þ%, respectively. Since these numbers are already very low, the probability for the atoms to be in levels further away from the circular is negligible. We thus estimate a preparation fidelity on the order of 99.5%.

APPENDIX E: rf SETUP
We generate the two quadratures of the radio-frequency fields by using two sets of four synthesizers with a global phase difference of π=2. For each set, we optimize the relative phases and amplitudes in order to generate a purely σ þ -polarized rf field with the four ring electrodes. We control the field global amplitude of each quadrature by applying control voltages V r ðtÞ and V i ðtÞ to sets of mixers used as voltage-controlled attenuators. Control voltages are generated by a two-channel arbitrary waveform generator with a 1 ns time resolution. (Fig. 11).
The field F rf ðtÞ is the sum of the two quadrature vector fields defined by F r ðtÞ ¼ F r ðtÞ½cosðω rf tÞu x þ sinðω rf tÞu y ; In order to determine the conversion between F r ðtÞ and V r ðtÞ [or F i ðtÞ and V i ðtÞ], we first record how the amplitudes of the rf signals after the amplifiers, S r k and S i k , vary as we change the control voltages V r and V i . Figure 12 presents the results of this measurement (normalized to the value of the amplitude for V r ¼ V i ¼ V 0 ¼ 1.52 V). We observe that all channels behave quite similarly (up to a ∼10% deviation for channel 2 and large values of the drive voltage). Using the measurement of the k ¼ 3 channel, we define ). The resulting signals are combined two by two using 3-dB couplers and sent to amplifiers, which feed the four ring electrodes surrounding the experiment (in gray).
We then use the atomic signal to measure the amplitude of the rf field F 0 r corresponding to V r ¼ V 0 . To that end, we prepare the atoms in the 52c state and apply to the mixers a pulse V r ðtÞ of constant amplitude V 0 and variable duration τ. Such a rf pulse induces a rotation of the spin coherent state describing the atoms at an angular frequency Ω 52 rf , which is directly proportional to the amplitude F 0 r of the rf field at the position of the atoms when V r ðtÞ is equal to V 0 . By measuring the population of the levels 52c, 52e 1 , and 52e 2 as a function of τ, we infer Ω 52 rf and measure F 0 r ¼ 4.37 AE 0.03 V=m. We measure with the same method the amplitude F 0 i ¼ 4.53 AE 0.02 V=m of the rf field when V i ðtÞ ¼ V 0 .
We finally have We invert these equations to convert the optimal pulse amplitude F r ðtÞ and F i ðtÞ into the time-dependent voltages V r ðtÞ and V i ðtÞ. This method assumes that all mixers provide exactly the same attenuation, and neglects the frequency response of f r ðVÞ and f i ðVÞ as well as the nonlinearity of the amplifiers at the end of the rf generation chain. It thus only provides a rough calibration. After the conversion, we perform a final "closed loop" optimization of the shape of V r ðtÞ and V i ðtÞ on the atomic signal as described in Appendix F.

APPENDIX F: EXPERIMENTAL OPTIMIZATION
OF THE PULSES

Optimization of the 52c preparation
We optimize the first 40 ns of the pulse by globally rescaling the amplitude of the pulse by an adjustable factor λ. This provides a first-order compensation for the uncertainty on the determination of f r , f i , F 0 r , and F 0 i and on the frequency response of the mixer. The last 73 ns of the pulse correspond to a rotation of the spin coherent state toward the north pole of the Bloch sphere. This part is very sensitive to phase and amplitude calibration of the rf. For a given value of λ, we directly optimize iteratively the amplitude of the plateaus of V r ðtÞ and V i ðtÞ on the population of 52c.
We repeat the optimization for different values of λ between 0.9 and 1. Figure 13 presents the populations P c , P e 1 , and P e 2 for each value of λ once the plateaus have been optimized. The flat optimum shows that the optimized pulse is very robust to small changes. We finally choose the value λ ¼ 0.95, which has the highest P c and the lowest P e 2 .

Optimization of coherence test rf pulse
For demonstrating the coherence of the optimal circularization pulse, we start the sequence with a superposition of levels 52c and 50c. We have to adapt the optimal pulse in order to additionally preserve the population of 50c. In the static field F 0 , the Stark frequency in the n ¼ 50 manifold is only Δ=2π ¼ 10 MHz smaller than ω rf =2π. Therefore, the rf pulse also affects the state of the atoms in the n ¼ 50 manifold. The first 40 ns of the optimal pulse transfer 50c into a spin coherent state jθ; ϕi 50 . We optimize the last  73 ns of the pulse so that they induce a rotation that maps jθ; ϕi 50 into 50c.
In the frame rotating at frequency ω rf =2π, a rf field of fixed phase and amplitude induces a rotation of the effective spin defined by the rotation vector, where ϕ is the phase of the pulse and Ω 52 rf the Rabi frequency, proportional to the rf amplitude. In the n ¼ 50 manifold, due to the detuning Δ, the same pulse induces a rotation of the effective spin defined by where Ω 50 rf ¼ ð50=52ÞΩ 52 rf . To ensure that jθ; ϕi 50 is rotated to 50c at the end of the pulse, we replace the single 73 ns pulse of constant amplitude (phase ϕ, Rabi frequency Ω 52 rf;0 ) by two pulses of duration τ 1 and τ 2 , with same phase ϕ but different amplitudes Ω 52 rf;1 and Ω 52 rf;2 , such that The condition (F1) ensures that the two-amplitude pulse induces in the n ¼ 52 manifold the same rotation as the constant amplitude pulse. By tuning the ratio between Ω 52 rf;1 and Ω 52 rf;2 and τ 1 and τ 2 , it is possible to ensure at the same time that the part of the wave function in the n ¼ 50 manifold returns into 50c at the end of the pulse. We optimize experimentally fΩ 52 rf;1 ; Ω 52 rf;2 ; τ 1 ; τ 2 g, with the constraint (F1) until the pulse transfers 50c into 50c [measured efficiency 98.2(5)%].
We implement the time reversal of the preparation pulse [first 113 ns of Fig. 3(a) in the main paper] by inverting the amplitude of one of the quadratures and programming the pulse backward in time. Experimentally, we fine-tune again the values of fΩ 052 rf;1 ; Ω 052 rf;2 ; τ 0 1 ; τ 0 2 g of the time-reversed pulse to ensure that it transfers 50c into 50c [measured efficiency 98.2(8)%] and 52c into 52m 2 [measured efficiency 88.1(9)%].
3. Optimization of the pulse preparing the superposition 52m 1 and 52c The first part of the pulse produces a superposition of a spin coherent state and of the rf dressed state that adiabatically evolves into 52m 1 when the rf is turned off. The second part of the pulse brings the spin coherent state to 52c, and slowly switches off so that population of the dressed state ends up in 52m 1 .
The optimization process is similar to that of the circularization pulse. First, we globally rescale the first 80 ns of the pulse with a scaling factor (leading to λ ¼ 0.925). Then, we adjust the amplitude of the two quadratures during the last part of the pulse to optimize the transfer of the spin coherent state into the circular state. To ensure that the other part of the wave function adiabatically follows the instantaneous dressed state of the rf, we vary linearly V r ðtÞ and V i ðtÞ between t ¼ 80 ns, t ¼ t 1 ¼ 136 ns, and t ¼ 167 ns. By varying the values of V r ðt 1 Þ and V i ðt 1 Þ characteristic of the linear ramps, we optimize the probability to end up in 52c, without affecting the probability to end up in 52m 1 at the end of the pulse. Note that the final ramp is slightly longer than in the simulation to compensate for the shape of f r and f i (Fig. 12): the rf field amplitudes F r ðtÞ and F i ðtÞ go to zero faster than the amplitude of V r ðtÞ and V i ðtÞ.
However, this optimization was not sufficient to obtain a perfect balance between 52c and 52m 1 . We found that the final population of 52m 1 is very sensitive to the amplitude of the optimal pulse during its first 4 ns. At this timescale, we can no longer correct the frequency response of the mixer by a global scaling factor λ. We need to rescale independently the amplitude of the driving voltage [V r ðtÞ → αV r ðtÞ and V i ðtÞ → αV i ðtÞ] for t < 4 ns. Figure 14 presents the final population of 52m 1 as a function of the scaling factor α. For α ¼ 2.8, we obtain a final population of P m 1 ¼ 0.502ð6Þ. 4. Optimization of the time-reversed pulse recombining the 52m 1 and 52c To measure the coherence of the superposition of 52m 1 and 52c, we program the time reversal of the pulse of Fig. 4(a). In principle, as the pulse of Fig. 4(a) performs the transfer and the evolution is unitary, the time-reversed pulse must perform the transfer j52ci → 1 ffiffi ffi 2 p ð−j52m 2 i þ j52m 2 i ⊥ Þ; where j52m 2 i ⊥ is a state orthogonal to j52m 2 i. Therefore, as we vary the relative phase of the superposition of j52ci and j52m 1 i before applying the time-reversed pulse, one should observe constructive and destructive interferences in the amplitude of probability to end up in j52m 2 i. The visibility of this interference pattern gives the degree of coherence of the superposition. Experimentally, we find that if we simply program V r ðtÞ → −V r ð−tÞ and V i ðtÞ → V i ð−tÞ, the resulting pulse transfers the state 52c into 52m 2 with a probability p c→2 ¼ 0.61ð2Þ and the state 52m 1 into 52m 2 with a probability p 1→2 ¼ 0.28ð1Þ. This would lead to a strong limitation of the visibility of the interference. The remaining population is found to be in 52m 1 [p c→1 ¼ 0.25ð1Þ for 52c and in p 1→1 ¼ 0.56ð2Þ for 52m 1 ]. To optimize the contrast of the fringes, we add after the time-reversed pulse an additional rf pulse, tuned on the 52m 1 − 52m 2 transition, whose phase is chosen to decrease p c→2 and increase p 1→2 . With this additional pulse, we obtain p final c→2 ¼ p final 1→2 ≈ 0.44.

APPENDIX G: ELECTRIC FIELD NOISE AND EXTRAPOLATION TO ZERO FIELD
Because of the differential Stark effect between the levels 52m 2 , 52m 1 , and 52c, the visibility of the fringes of Figs. 3(b) and 4(b) is very sensitive to the electric field noise. To quantify this effect, we vary the time delay between the two rf pulses that prepare and recombine the "cat states" (j50ci þ j52m 2 i or j52ci − j52m 1 i, respectively) and measure, for each delay, the visibility of the fringes. The longer the atoms stay in the superposition, the more sensitive the relative phase is to the electric field, reducing the visibility of the fringes. To calibrate the sensitivity of the relative phase to the electric field, we record the fringes for two slightly different values of the static electric field [4]. Figure 15 shows the visibility as a function of the phase sensitivity ξ [in rad/(mV/m)] for the two superpositions. Assuming a Gaussian electric field noise, we expect the visibility V to be where σ F is the variance of the electric field noise and V 0 the intrinsic visibility in the absence of noise. A fit to the experimental data gives V 0 ¼ 1.00ð1Þ [and σ F ¼ 13.4ð1Þ mV=m] for the superposition j50ci þ j52m 2 i. For the superposition j52ci þ j52m 1 i, we find V 0 ¼ 0.97ð2Þ and a slightly cleaner field σ F ¼ 11.6ð1Þ mV=m (the two sets of data have been recorded at a few months time interval). By symmetry, the phase sensitivity of the preparation of the superposition is half that of the full process of preparation and recombination. As a result, the degree of coherence C of the superposition j52ci − j52m 1 i after the rf pulse is where ξ 0 is the phase sensitivity for the shortest Δt. The final fidelity F is then ffiffiffiffiffiffiffiffiffiffiffiffi ffi P c P m 1 p Þ ≥ 0.93: Visibility of the fringes measuring the coherence of (a) j50ciþ j52m 2 i and (b) j52ci þ j52m 1 i, as function of the phase sensitivity of the fringes to the electric field ξ. The points are experimental (with statistical error bars), the line is a Gaussian fit of the form V ¼ V 0 expð−ξ 2 σ 2 F =2Þ, where V 0 is the intrinsic visibility that the fringes would have in the absence of electric field noise.