WorksPhysics and Astronomy
Gradient Ascent Pulse Engineering for Multi-Time Correlations
Abstract
In quantum optimal control, gradient methods are often used to optimize functions of a state or propagator at a final time, while experimentally addressable observables and performance metrics may depend on multi-time correlations. We develop QRT-GRAPE, which combines the quantum regression theorem (QRT) with Gradient Ascent Pulse Engineering (GRAPE) to obtain analytic gradients for objective functions defined by multi-time correlations. The method incorporates the time evolution of an auxiliary operator in QRT into the forward propagation and backward adjoint propagation used in GRAPE. As an application, we consider the trade-off between the photon-emission probability and the conditional Bell-state fidelity in heralded remote entanglement generation among two nonidentical atom–cavity systems. Starting from a flat-top Rabi pulse pair selected by a search over pulse amplitudes and durations, QRT-GRAPE increases the photon-emission probability product from to and the conditional Bell-state fidelity from to .
Introduction
The correlation functions allow us to analyze statistical relations between quantities and events that do not necessarily occur in the same times. They are widely used in different disciplines, especially in which the relations and dynamics are determined by statistical laws. In quantum physics, they become fundamental to connect quantum dynamics and experimentally accessible observables in different times and to calculate operationally important functions such as fidelity between quantum states or quantum entanglement. They define spectroscopic and linear-response signals [1, 2], determine dynamical structure factors used to characterize quantum many-body systems [3, 4], and play a role in principal component analysis of continuous variable quantum states [5, 6].
In quantum optics, the correlation functions characterize single-photon sources. The second-order coherence function, , measures photon statistics and antibunching [7, 8, 9, 10, 11], whereas the unnormalized first-order coherence function, , captures temporal coherence and temporal mode structure. Temporal mode overlap is central to two-photon interference and photon indistinguishability [11, 12, 13, 14], which directly characterize fidelities of atom–photon entanglements [15, 16] and remote entangled states generated by interferometric measurements [14, 17, 18, 19, 20, 21, 22, 23, 24].
Since the states evaluated by correlation functions form fundamental building blocks in quantum information processing, the correlation functions are directly related to the performance of protocols of quantum technology. For example, photonic links in quantum networks and distributed quantum architectures rely on the correlation properties of propagating photons [25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35].
Moreover, correlation-based performance quantities also arise in distributed quantum sensing based on entanglement [36, 37].
Specific values of correlation functions may be desired between selected observables in particular times. As such, a correlation function becomes the figure of merit in designing controlled dynamics leading the system from the initial state to the target configuration. This is the subject of optimal control theory. In this case, its goal is to optimize control pulses using objective functions defined by correlation functions.
Quantum optimal control (QOC) provides multiple methods for designing time-dependent control pulses to achieve desired dynamical objectives [38, 39]. Among them, Gradient Ascent Pulse Engineering (GRAPE) [40] is a widely used powerful approach. GRAPE and its variants have been applied to state preparation [41, 42, 43], gate optimization [40, 44, 45], quantum sensing [46, 47], and open-system control [48, 49, 50]. However, as shown in Figure 1, standard GRAPE is primarily formulated for objective functions evaluated at a single final time, such as state or gate fidelity. In contrast, correlation functions depend on multi-time observables, to which standard GRAPE methods are not applicable. The goal of this work is to adapt the GRAPE method to correlation function targets.
Figure 1. Conceptual comparison between GRAPE for objective functions depending on the state or propagator at the final time and QRT-GRAPE for objective functions constructed from correlation functions. In GRAPE, the state is propagated forward in time and the adjoint variable backward in time, without auxiliary propagation initiated by an operator insertion. In QRT-GRAPE, the operator insertion required by the quantum regression theorem starts auxiliary propagation for pairs of times . This construction accommodates objective functions constructed from correlation functions, including measures of photon indistinguishability and temporal mode matching in open quantum systems.
Multi-time correlations such as a first-order coherence function [51, 52, 53] are the subject of quantum regression theorem (QRT). For objective functions constructed from multi-time correlations, the strategy shown in Figure 1 includes an additional step beyond the standard time evolution: an operator corresponding to the correlator is inserted at the earlier time, and the resulting auxiliary operator is then propagated to evaluate the correlation function. Therefore, the standard GRAPE is not directly applicable to these correlation functions.
In this paper, we develop QRT-GRAPE, which combines QRT with GRAPE to optimize multi-time correlation functions. When the requirements of the QRT are satisfied, the auxiliary operator evolves with the same time-dependent generator as the density operator. We use the common time-dependent generator governing the density operator and the QRT auxiliary operator to incorporate the operator insertion and auxiliary propagation into GRAPE. This yields analytic gradients for objective functions constructed from multi-time correlations and opens the GRAPE-type methods to the new class of objective functions.
As a proof-of-concept application, we optimize two atom–cavity single-photon sources for heralded remote entanglement generation, which has been demonstrated experimentally [21]. For distributed quantum information processing, both the entanglement generation probability and fidelity are crucial metrics, but they are subject to a trade-off [14]. Starting from Rabi pulse pairs selected by parameter sweeps, QRT-GRAPE optimizes the Rabi frequencies and increases both the photon-emission probability product and the conditional Bell-state fidelity. This result indicates that our method can overcome the fundamental bottleneck in quantum networks, namely the trade-off between the photon-emission probability and the conditional Bell-state fidelity. In practical networks, the parameters of the two atom–cavity systems can differ [35, 54]. We therefore investigate the case of connecting two nonidentical cavities, in which we have swept the atom–cavity coupling ratio and the external cavity decay rate ratio . At all pairs of ratios, QRT-GRAPE increased the objective function value , the photon-emission probability product , and the Bell-state fidelity relative to the values obtained with the Gaussian Rabi pulse pair selected at the same pair of ratios. In this sweep, the results suggest that weak atom–cavity coupling limits the optimized performance more strongly than the external cavity decay rate mismatch.
This paper is organized as follows. Section II reviews GRAPE, the Lindblad master equation, and the QRT. § 1 derives the QRT-GRAPE gradient formula. Section IV presents the cavity QED application and numerical results. § 2 concludes with a discussion of broader applicability and future directions.
Preliminaries
GRAPE
GRAPE [40] is a widely used gradient-based quantum-control method. We summarize the forward–backward structure needed for the QRT-GRAPE extension, following Ref. [55]. For simplicity, we explain the case of pure states, while it can be easily extended for the mixed state with vectorization of matrices, as we describe later.
We consider a quantum system governed by the Hamiltonian
where is the static Hamiltonian and is the control Hamiltonian associated with the real control function . GRAPE partitions the total evolution time into slices of duration and takes
where . With , the th sliced propagator is , so that the final state is
For a real-valued objective function at the final time, with , we define the terminal adjoint state as
The derivative is taken with held fixed. Since is real, the two conjugate contributions to its first-order variation give
where is the first-order change in the terminal state induced by variations of the control values. For a change in a single control value, the first-order change in is . Since only depends on this control value, Eq. (3) gives
Substituting this expression into the first-order variation of and using gives
where and are given by forward and backward propagations:
respectively. The derivative is evaluated by
For sufficiently small , the slice derivative reduces to the usual GRAPE approximation,
Substituting this relation gives the gradient:
The partial derivatives of the objective function for all can therefore be evaluated from the states and adjoint states obtained in one forward and one backward propagation, respectively. Conversely, the naïve finite-difference method requires to solve the full propagations iteratively for each partial derivative. Thus, the required number of propagations scales as for finite differences, whereas GRAPE requires only two: one forward and one backward propagation.
Quantum Regression Theorem
For a Markovian open quantum system, the dynamics are governed by the Lindblad master equation [56, 57]:
where is the Liouvillian superoperator, denotes the system Hamiltonian, and the Lindblad operators describe dissipation.
Expectation values are given by , while the density matrix cannot determine multi-time correlations such as the first-order correlation . Computing such correlations therefore requires an additional relation.
The quantum regression theorem (QRT), introduced by Lax [51], provides a way to obtain multi-time correlations through auxiliary propagation. For simplicity, we first consider a two-time correlation and later generalize the construction to multi-time correlations. For , we consider the correlator . According to QRT, the correlator is represented as
where the auxiliary operator evolves from to with the same generator as the Lindblad master equation,
For the oppositely ordered correlator, , the insertion of gives another auxiliary operator , and the correlator is evaluated with at :
The subsequent auxiliary propagation uses the same time-dependent Liouvillian as in (14).
A time-ordered multi-time correlator,
is obtained by recursively applying the QRT. For instance, we define the th auxiliary operator,
followed by the auxiliary propagation according to Eq. (14). The resulting correlator is given by
The same procedure can be used for an opposite ordering,
and a symmetric time-ordered correlator,
with a slight modification of the auxiliary operator.
QRT-GRAPE
We formulate QRT-GRAPE for real-valued objective functions constructed from two-time correlations on the chronological ordering . Correlators with use the distinct QRT construction in Eq. (15). We first vectorize the open-system dynamics and QRT construction, then derive the analytic gradient. This procedure can be easily generalized to multi-time correlators. Figure 1 conceptually compares GRAPE and QRT-GRAPE, and Figure 2 summarizes the forward, auxiliary, and adjoint propagations used below.
Figure 2. Computation graph for the QRT-GRAPE gradient, illustrated for the auxiliary region case . The upper row gives the ordinary forward propagation, the middle row gives the QRT auxiliary-branch initiated by the operator insertion at , and the lower row gives propagation of the adjoint variable backward in time from the evaluation of the correlator with at . The derivative of with respect to is obtained by evaluating . The definitions of and for all positions of interval are given in Eqs. (33) and (34), respectively.
Superoperator Formalism
Section II B defines the insertion of the operator and auxiliary evolution. We refer to the auxiliary propagation initiated by this insertion as the QRT branch. After vectorization, all procedures in QRT are analogous to those in the state-vector representation. This vectorized representation allows us to combine the QRT with the backward adjoint propagation used in GRAPE.
Given an orthonormal basis , we vectorize an operator according to
Under this representation, left and right multiplication correspond to the following Kronecker products:
More generally, . We use this identity for the Lindblad generator and operator insertion below.
Applying the identity to the commutator gives
and the Lindblad equation becomes , where denotes the Liouvillian in this formalism.
The initial condition in QRT is vectorized as
By defining the unnormalized maximally entangled vector,
the trace takes the inner-product form:
Therefore, the two-time correlator is
Gradient of Two-Time Correlators
We represent each control function by piecewise constant values on interval . The Liouvillian is differentiable with respect to these control values. We collect the control values in :
For a single operator pair , we define the two-time correlator for by
Using the QRT, we express this correlator in terms of the propagators as
where with discretized Lindblad generator , and denotes the initial state. For , Eq. (15) instead uses the earlier-time insertion , propagation to , and readout with .
Differentiating the correlator with respect to the control value only affects the sliced propagator . With the corresponding auxiliary and adjoint states, the correlator derivative takes the compact form:
where the vectorized operator is
and the adjoint variable for the correlator is
Using Eq. (9), the derivative of the th propagator is
For sufficiently small , the derivative reduces to
In the numerical calculations, we evaluate the exact derivative in Eq. (35) using the Fréchet derivative of the matrix exponential.
Gradient of objective functions constructed from correlations
In Sec. III B, we derived the derivative of a single two-time correlator. To construct objective functions from several correlation functions, we consider the set of two-time correlators associated with the operator pairs and define the general objective function by
where is a differentiable real-valued function and is an optional penalty term. The control values are real and the correlators are generally complex, so we use Wirtinger derivatives [58, 59] and , defined for by treating and as independent variables, which gives
For real and real control values, the conjugate terms combine into the real-valued form:
where
defines the sensitivity coefficients . Combining Eqs. (39) and (32) gives the QRT-GRAPE gradient,
where and are the auxiliary and adjoint states associated with the operator pair . The gradient is therefore a linear combination of the derivatives of the two-time correlators, even when is a non-linear function.
For objective functions constructed from multi-time correlations, the two-time differentiation procedure can be combined with the recursive multi-time QRT procedure described in Sec. II B to extend QRT-GRAPE beyond the two-time case.
QRT-GRAPE Optimization of Cavity-QED Photon Sources
In the previous section, we established QRT-GRAPE. In this section, we apply it to heralded remote entanglement generation between two atom–cavity single-photon sources, which has been demonstrated experimentally [21] and analyzed theoretically [14]. Figure 3(a) shows the atom–cavity system, and Figure 3(b) shows the entanglement-swapping protocol, in which photons emitted from two such systems interfere at a beam splitter and are detected. Here, we use QRT-GRAPE to optimize the Rabi frequencies driving the two systems.
Figure 3. Cavity-QED setting and scope of the photon-source demonstration. (a) The numerical model directly propagates each three-level -type atom–cavity module. The pump drives the transition, while denotes the controllable diagonal detuning of in the rotating-frame Hamiltonian. The atom–cavity coupling connects and , and denotes the cavity detuning. The cavity photon leaves through the desired cavity output, characterized by , or is lost internally, characterized by . Spontaneous emission from either returns population to , characterized by , or transfers it to the zero-excitation state , characterized by . (b) The broader entanglement-swapping protocol motivates the objective function for temporal mode overlap: photons emitted from two independently driven atom–cavity modules interfere at a beam splitter and are detected in a Bell-state measurement.
Atom–Cavity System
{Figure 3(a)} shows a three-level -type atom with ground states and and excited state , whose energies are (). The transition couples to a single-mode cavity of frequency with coupling . An external laser with Rabi frequency drives the transition .
Since the total excitation number is at most one, the coherent dynamics are restricted to , where denotes atomic state and cavity Fock state . In the basis , the
rotating-frame Hamiltonian takes the form
where denotes a cavity detuning, and denotes an atomic detuning with instantaneous laser frequency .
Spontaneous decay rate from back to and to are characterized by and , respectively, while cavity decay from is characterized by and for cavity output and internal cavity loss. To preserve trace, we include the zero-excitation state to the Hilbert space:
Here, denotes the zero-excitation state reached after cavity decay or spontaneous emission by . On this four-state space, , and the dynamics are governed by the Lindblad master equation,
where the index labels the Lindblad operators defined below:
Here, is the amplitude decay rate of due to spontaneous emission, so the corresponding population decay rate is . Hereafter, for the cavity-QED example and its numerical appendices, we express frequencies (including detunings), coupling strengths, decay rates, and Rabi frequencies in units of , and times in units of . We use the same symbols for the resulting dimensionless quantities.
The returns population to , allowing subsequent re-excitation, whereas transfers population to without contributing to the desired cavity output. Both cavity output and internal cavity loss transfer population from to , with only counted as useful output.
Objective functions for remote entanglement generation
We apply QRT-GRAPE to atom–photon entanglement generation for cavity-based entanglement swapping, as illustrated in Figure 3(b). The emitted photons interfere at the central node and are subsequently measured by single-photon detectors. The success probability and fidelity of atom-atom entanglement are determined by the generation probability of atom–photon entanglement and the first-order coherence of each photon, respectively. Therefore, the numerical model treats the two independent atom–cavity sources and evaluates their output-field correlations and photon-emission probabilities. The noises outside the atom-cavity systems are neglected since they are not relevant to the optimization problem.
We consider two atom–cavity systems A and B. For the system , the photon-emission probability is obtained from the output photon flux as
We quantify the temporal mode indistinguishability of the two emitted photons by the normalized wave-packet overlap :
When the temporal mode mismatch is the only noise source, determines the conditional Bell-state fidelity [14]:
Maximizing the real part of improves the conditional fidelity but does not reward photon-generation probability. To combine photon generation with temporal mode overlap in a single optimization target, we use the unnormalized overlap [14]:
The prefactors follow from the input-output relation . For each source,
so exchanging and maps the correlation to its complex conjugate and makes real.
Hereafter, we discretize the objective function for numerical optimization. Let denote the th time point at which the two-time correlation functions are evaluated, and let denote the trapezoidal integration weight at . The discretized objective function is given by Eq. (51).
Writing , we use the Hermitian symmetry to obtain the upper-triangular form:
The trapezoidal weights are given by
The objective function is evaluated using the correlation functions at the time points . The subsequent time is the final edge of the last simulation interval and is not itself a correlation calculation point.
For source , we decompose the th-interval Hamiltonian into the static term and two control Hamiltonians:
where and are the constant values of the Rabi frequency and atomic detuning on . The corresponding control Hamiltonians are
while contains the cavity detuning and atom–cavity coupling.
Because the two sources evolve independently, differentiating with respect to acts only on the correlator for the source A. Then, the correlator for the source B, , is treated as a scalar weight, and vice versa.
Since the static Hamiltonian is independent of the control values, the derivative of the Liouvillian with respect to either control value is determined solely by the corresponding control Hamiltonian:
We evaluate the propagator derivative using Eq. (35), with the Liouvillian derivative in Eq. (56). For source A, Eq. (32) gives
Here denotes the vectorized auxiliary operator at interval , while denotes the adjoint state for the correlator , as given in Eqs. (33) and (34), respectively. For notational simplicity, we denote this correlator derivative by
Note that for since the correlator is independent of the control after .
For source A, the off-diagonal terms in Eq. (52) give for . For the real diagonal correlators in this application, we extend the objective to depend on through , giving . Substitution into Eq. (39) gives the gradient with respect to :
The same expression applies to source B by changing and .
Optimization results of entanglement swapping
The parameters used in the examples are listed in Table 1. For the comparison in Table 2, we optimize the Rabi frequencies and with both atomic detunings fixed at zero.
| Parameter | Cavity A | Cavity B |
| 1.4 | 1.0 | |
| 0 | 0 | |
| 0.9 | 1.1 | |
| 0.08 | 0.08 | |
| 0.3 | 0.3 | |
| 0.7 | 0.7 |
Table 1. Physical parameters used for the two atom–cavity sources in the numerical demonstration. All frequencies and decay rates are normalized by .
| Initialization | Stage | ||||
| Gaussian | Initial | 0.2354 | 0.9401 | 0.5957 | 0.4489 |
| Optimized | 0.2857 | 0.9441 | 0.6569 | 0.4896 | |
| Flat-top | Initial | 0.2441 | 0.9356 | 0.6065 | 0.4620 |
| Optimized | 0.2857 | 0.9441 | 0.6569 | 0.4896 |
Table 2. Comparison of the objective function value , conditional Bell-state fidelity , and external photon-emission probabilities and before and after QRT-GRAPE optimization, for the physical parameters in Table I. We optimize the discretized Rabi frequencies and while fixing both atomic detunings at zero. Fixing the first Rabi frequency of each source at zero leaves 198 adjustable Rabi frequencies.
We divide the Rabi pulse driving period for sources A and B into 100 equal time intervals, with one
constant Rabi frequency for each source in each interval. The Rabi frequency of each source in the first interval is fixed at zero. QRT-GRAPE optimizes the Rabi frequencies in the remaining 99 intervals of each source, giving 198 adjustable Rabi frequencies in total. Both atomic detunings remain fixed at zero. Each adjustable Rabi frequency is bounded by . Both Rabi frequencies and both atomic detunings are zero during an additional decay interval, allowing the remaining cavity photon to decay. We use box-constrained L-BFGS-B with the QRT-GRAPE gradient. Appendix A gives the further numerical settings.
We considered Gaussian and flat-top Rabi pulses for comparison to QRT-GRAPE. For the Gaussian Rabi pulses, we vary the pulse amplitudes and widths. For the flat-top Rabi pulses, we vary the pulse amplitudes and durations (see details in Appendix A). For each pulse shape, we select the A/B pulse pair with the highest objective function value among the candidate pairs. Starting from each selected pulse pair, QRT-GRAPE optimizes 99 Rabi frequencies per source, with both atomic detunings fixed at zero. Before QRT-GRAPE optimization, the selected pairs of Gaussian and flat-top Rabi pulses gave objective function values and 0.2441, respectively. Starting from either selected pulse pair, QRT-GRAPE reached , corresponding to relative increases of 21.4% and 17.0% for the Gaussian and flat-top pulses, respectively. Table 2 summarizes the corresponding objective function values , conditional Bell-state fidelity, and external photon-emission probabilities before and after optimization.
Appendix A 2 shows how the objective function value and the norm of its gradient with respect to the adjustable Rabi frequency change during optimization starting from the selected Gaussian and flat-top Rabi pulse pairs. We also performed 100 QRT-GRAPE optimizations starting from independently generated random Rabi pulses, with both atomic detunings fixed at zero. These optimizations are described in Appendix A 3. Of the 100 optimizations, 87 ended with a final objective function value of , showing that high objective function values were reached from many different initial Rabi pulse pairs. Some optimizations terminated at lower values. These results suggest that running the optimization from several initial Rabi pulse pairs and retaining the largest final value of can help reduce dependence on initialization in this example.
We also tested whether penalizing differences between neighboring Rabi frequencies reduces steep variations in the optimized pulses. Starting from the Gaussian Rabi pulse pair selected for Table II, we optimized only the Rabi frequencies, with both atomic detunings fixed at zero (Appendix B). Penalizing differences between neighboring Rabi frequency values produced smoother optimized pulses, with an approximately 0.24% decrease in for relative to optimization without the penalty. In this example, QRT-GRAPE therefore found smoother pulses while retaining nearly the same objective function value. The roughness measure used in this test excludes the change in each Rabi frequency when the drive is turned off. We expect that including this change in the penalty would also reduce the magnitude of the jump at the end of the drive.
So far, we have considered the nonidentical atom–cavity systems with the fixed parameters in Table I. In practical networks, the parameters of the two atom–cavity systems can differ [35, 54]. We therefore investigate how our method improves the objective function value for various combinations of and . In the simulation, we use the parameters of system A in Table I except for , while the and are swept such that . Since the photon-emission probability depends on the cavity decay rate , we choose the optimal cavity-decay rate for the cavity A, which maximizes the photon-emission probability [60]:
Substituting the parameters given in Table I, the optimal decay rate becomes 0.48. In Figure 4, the atom–cavity coupling ratio and external cavity decay rate ratio each take seven equally spaced values from 0.5 to 1.0, giving a grid. At each grid point, we construct Gaussian Rabi pulses for each source from 20 amplitudes in and 20 widths in , with and pulse center . Appendix A defines the Gaussian Rabi pulses and describes how they are discretized into 100 piecewise-constant time intervals. Each source has 400 candidate pulses; we evaluate the objective function for all pairs of Gaussian pulses from sources A and B. The pair with the highest value of is denoted by . At each grid point, we run QRT-GRAPE once from that selected pair, optimizing the Rabi pulses while keeping both atomic detunings fixed at zero.
Figure 4. Objective function value, product of photon-emission probabilities, and Bell-state fidelity after QRT-GRAPE optimization as a function of the parameters of source B. The horizontal and vertical axes are and , respectively, with and . (a–c) Optimized objective function value , photon-emission probability product , and conditional Bell-state fidelity , respectively. (d,e) Relative gains in percent for and , respectively. (f) Dimensionless absolute difference . The subscript G denotes the corresponding value for the Gaussian Rabi pulse pair selected at the same grid point. All three quantities increase at every grid point relative to their values for that Gaussian Rabi pulse pair.
Figure 4(a) shows the optimized objective function value . Figure 4(b) and Figure 4(c) show the corresponding optimized photon-emission probability product and conditional Bell-state fidelity , respectively. The upper-right grid point corresponds to identical atom–cavity sources. With , reducing from 1 to 0.5 decreases by 40.2%. With , reducing from 1 to 0.5 decreases by 10.9%.
Figure 4(d) and Figure 4(e) show the relative gains in the objective function value and photon-emission probability product , respectively. The relative gain in ranges from 34.8% to 86.0%, with a median of 52.5%. The relative gain in ranges from 33.8% to 73.2%, with a median of 48.9%. Figure 4(f) shows the absolute Bell-state fidelity difference , which ranges from 0.0032 to 0.0322, with a median of 0.0098. All three increases are positive at all 49 grid points. Every comparison uses the Rabi pulses optimized at a particular grid point and the Gaussian Rabi pulses selected at that same grid point.
Reducing the atom–cavity coupling lowers the internal cooperativity and the upper bound on the single-photon generation efficiency [60]. This is consistent with the substantial decrease in the objective function value at weaker coupling even after optimization of the Rabi pulses. The external cavity decay rate mismatch produces a smaller decrease in the optimized objective function value than the reduction in coupling strength. Over the parameter range studied here, these results suggest that weak atom–cavity coupling is the main factor limiting the optimized performance, rather than the external cavity decay rate mismatch.
We also compared optimization of the Rabi frequencies alone with joint optimization of the Rabi frequencies and atomic detunings. Starting from the Gaussian Rabi pulse pair selected for Table II, we performed 30 joint optimizations with independently generated initial atomic detuning values (Appendix C). None of the 30 optimizations of both Rabi frequencies and atomic detunings exceeded the existing result obtained by optimizing the Rabi frequencies alone.
We verified the analytic gradient by comparing it with centered finite differences with respect to selected adjustable values of the Rabi frequencies and atomic detunings of both sources. The analytic derivatives of agreed with the centered-finite-difference estimates of the corresponding partial derivatives of to high numerical accuracy, supporting the derivative construction and its numerical implementation. In the timing comparison, evaluating and its analytic gradient was 9.40 times as fast as computing the full gradient by centered finite differences. Further details of the validation are provided in Appendix D.
Conclusion
In this paper, we developed QRT-GRAPE for optimizing objective functions constructed from multi-time correlation functions. The method incorporates the operator insertion and auxiliary propagation in QRT into the forward propagation and backward adjoint propagation used in GRAPE. This construction provides analytic gradients for objective functions defined directly by multi-time correlation functions, extending gradient-based optimal control beyond objective functions represented by the state at a single final time.
As an application of QRT-GRAPE, we considered heralded remote entanglement generation using two nonidentical atom–cavity single-photon sources. Starting from the flat-top Rabi pulses, QRT-GRAPE increased the objective function value from 0.2441 to 0.2857, corresponding to a relative improvement of 17.0%. In the same comparison, the photon-emission probability product increased from 0.2802 to 0.3216, while the Bell-state fidelity increased from 0.9356 to 0.9441.
We also varied the atom–cavity coupling ratio and external cavity decay rate ratio over a grid. At each of the 49 grid points, we selected the Gaussian Rabi pulse pair with the highest value of from the Gaussian pulse amplitude and width search and ran one QRT-GRAPE optimization starting from that pair. At every grid point, QRT-GRAPE increased the objective function value , photon-emission probability product, and Bell-state fidelity relative to the values obtained with the Gaussian Rabi pulse pair selected at the same grid point. The relative increase in ranged from 34.8% to 86.0%.
The present formulation assumes that, after each operator insertion required by the QRT construction, the auxiliary operator evolves under the same time-dependent generator that governs the density operator. Extending QRT-GRAPE to cases in which this QRT propagation rule is not valid will require a different treatment of multi-time correlations. A further practical challenge is the computational cost of evaluating correlations over many combinations of times, motivating more efficient propagation and selection of time points for evaluating multi-time correlation functions. Subject to this QRT condition, QRT-GRAPE provides a general framework for optimizing objective functions constructed from multi-time correlation functions.
Data Availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Acknowledgments
This work was supported by JST Moonshot R&D, Grant Number JPMJMS256E and JPMJMS256G, JST ASPIRE, Grant Number JPMJAP2427, JST COI–NEXT Grant No. JPMJPF2221, and Program for the Advancement of Next Generation Research Projects, Keio University. K.T. was supported by JST SPRING, Grant Number JPMJSP2123.
Optimization and initialization methods
In this Appendix, we describe the optimization settings and methods. Table III summarizes the settings for Table 2 and Figure 4. We divide into 100 intervals of width . For each Rabi frequency and atomic detuning, we use one value at the left endpoint of each interval and hold that value constant throughout the interval. The value of each function in the first time interval is fixed at zero. Both Rabi frequencies and both atomic detunings are zero during the additional decay interval from 10 to approximately 14.2373, which is divided into 10 intervals.
| Parameter | Value |
| Rabi-frequency bounds | |
| Atomic detunings | Fixed at zero |
| Number of adjustable Rabi frequencies | 198 |
| Initial Gaussian width | |
| Duration of the Rabi pulse driving period | 10 |
| Additional decay duration | |
| Number of time intervals in the Rabi pulse driving period | 100 |
| Decay intervals | 10 |
| Maximum iterations | 500 |
Table III. Numerical settings used for the two-source optimization and the cavity parameter study as shown in Table II. The Gaussian width range is used to select the initial Gaussian pulses. All time quantities are expressed in units of , where .
For Table 2 and Figure 4, we optimize the remaining 99 values of each Rabi frequency, and , within , giving 198 adjustable Rabi frequencies, while both atomic detunings remain fixed at zero. The optimizations from random initial Rabi pulses described below and the smoothness test each optimize 99 Rabi frequencies per source, with both atomic detunings fixed at zero, giving 198 adjustable values per optimization. For the detuning comparison in Appendix C, we reuse the result obtained by optimizing the Rabi frequencies with both atomic detunings fixed at zero, and perform 30 new optimizations with 396 adjustable control values each.
For Table 2, we first select pairs of Gaussian Rabi pulses and pairs of flat-top Rabi pulses by varying their amplitudes and Gaussian widths or flat-top durations with both atomic detunings fixed at zero, as described below. QRT-GRAPE then allows the Rabi frequency in each of the remaining 99 time intervals of each photon source to vary. The optimized pulses can therefore depart from Gaussian or flat-top shapes. We use box-constrained L-BFGS-B with the analytic QRT-GRAPE gradient of and at most 500 iterations.
Initial Rabi Pulse Pairs
We selected the initial Gaussian and flat-top Rabi pulse pairs for Table 2 as follows. The Rabi pulses were fixed at zero at and for . Both atomic detunings, and , were fixed at zero during the pulse parameter sweeps and the QRT-GRAPE optimizations reported in Table 2.
For Table 2 and Figure 4, the initial Gaussian Rabi pulse for source was obtained by discretizing the function
with , center , and . We evaluated at the left endpoint of each time interval and held that Rabi frequency constant over the interval. Both Rabi frequencies and both atomic detunings were zero during the additional decay interval.
The flat-top pulse is defined by
where and is the total flat-top pulse duration, including the rise and fall. The pulse is zero outside .
For each source in Table II, we constructed Gaussian Rabi pulses using 100 uniformly spaced amplitudes in and 100 uniformly spaced widths in . For the flat-top Rabi pulses, we used the 100 Rabi frequency amplitudes from the Gaussian pulse search and 100 uniformly spaced durations from 0.1 to 10. The Gaussian pulse search and the flat-top pulse search each had 10,000 candidate pulses per source and pairs of pulses from sources A and B. We evaluated the objective function for every Gaussian pulse pair and every flat-top pulse pair using the nonidentical source parameters in Table I. For each pulse shape, we selected the pair with the highest value of . Table IV lists the selected pulse amplitudes and widths or durations.
| Pulse shape | Source A | Source B |
| Gaussian | , | , |
| Flat-top | , | , |
Table IV. Parameters of the initial Rabi pulses selected for Table II. Here, is the Rabi frequency amplitude, is the Gaussian pulse width, and is the flat-top pulse duration for source . Values are rounded consistently with the main text.
Optimization Histories and Rabi Pulses
Figure 5 shows the QRT-GRAPE optimization histories starting from the Gaussian and flat-top Rabi pulse pairs selected for Table II.
Figure 5(a) shows that the two selected initial Rabi pulse pairs lead to nearly the same final objective function value. Starting from the Gaussian Rabi pulse pair, increases from 0.2354 to 0.2857 over 303 accepted iterations. Starting from the flat-top Rabi pulse pair, it increases from 0.2441 to 0.2857 over 186 accepted iterations. The optimized Rabi pulses also closely resemble one another for each source, as shown in Figure 5(c,Figure 5d). In both cases, the optimized Rabi frequencies increase through most of the driving period and reach the upper bound before .
Figure 5. QRT-GRAPE optimization histories and Rabi pulses for the Gaussian and flat-top pulse pairs used in Table II. (a) Objective function value . (b) Euclidean norm of the gradient of with respect to the 198 adjustable Rabi frequencies. The horizontal axis in (a,b) counts accepted L-BFGS-B iterations, with 0 denoting the initial pulses. (c,d) Initial and optimized Rabi pulses for sources A and B, respectively. Blue and orange indicate runs starting from Gaussian and flat-top Rabi pulses, respectively; dashed and solid lines in (c,d) denote initial and optimized pulses. Both atomic detunings are fixed at zero. The shaded region at is the additional decay interval, during which both Rabi frequencies are fixed at zero.
As shown in Figure 5(b), the Euclidean norm of the gradient of decreases overall, with fluctuations, but remains nonzero at termination. Several optimized Rabi frequencies reach the upper bound . Under these amplitude constraints, the Euclidean gradient norm need not vanish at an optimum because directions that would increase can be excluded by the bounds.
The gradual increase in the optimized Rabi frequencies produces pulses that are strongly asymmetric in time, as shown in Figure 5(c,Figure 5d). Ref. [14] indicates that asymmetric Gaussian pump pulses with a longer rise than fall could improve the conditional Bell-state fidelity at a given photon-emission probability in their model. The gradual rise found here may similarly promote population transfer from to while limiting the population of and the associated spontaneous emission. The switch to zero at is imposed by our control conditions.
Optimization from Random Initial Rabi Pulses
We also performed 100 QRT-GRAPE optimizations starting from independently generated random Rabi pulses, with both atomic detunings fixed at zero. The source pair is specified in Table I. For each source and each of the 99 adjustable time intervals, the initial Rabi frequency is picked independently from a uniform distribution on . These initial Rabi frequencies are used directly, without smoothing. QRT-GRAPE optimizes the 99 Rabi frequencies of each source within , giving 198 adjustable values in total.
Each trial uses L-BFGS-B with the analytic QRT-GRAPE gradient, at most 500 iterations, and . The trial stops earlier if either convergence criterion is met. We use the value of at termination in the histogram, including trials that reach the iteration limit. We count trials with an increase in separately from those for which L-BFGS-B reports convergence.
All 100 trials ended with a larger objective function value than their respective initial values. L-BFGS-B reported convergence for 79 trials, and the remaining 21 reached the 500-iteration limit. Figure 6 shows the distribution of the final values from all 100 trials. The final values ranged from 0.2226 to 0.2857, with a median of 0.2857. Of the 100 trials, 87 ended with .

Figure 6. Histogram of the objective function value at termination for 100 QRT-GRAPE optimizations starting from independent random Rabi pulse pairs, with both atomic detunings fixed at zero. The final values are grouped into 10 equal-width bins. The vertical axis is logarithmic, and the short marks along the horizontal axis indicate individual final values. Each trial runs for at most 500 L-BFGS-B iterations and stops earlier if a convergence criterion is met. Trials that reach the iteration limit are included.
The concentration of final values near in Figure 6 shows that high objective function values were reached from many different initial Rabi pulse pairs. The lower-valued tail shows that a single run can terminate with a smaller objective function value under the same stopping conditions. Running the optimization from several initial Rabi pulse pairs and retaining the largest final value of therefore provides a practical way to reduce dependence on initialization in this example.
Smoothing optimized pulses
The optimized Rabi pulses in Figure 5 contain steep variations between neighboring time intervals. To explore smoother pulses for experimental implementation while seeking high objective function values, we included a penalty on these variations during optimization.
All three optimizations started from the Gaussian Rabi pulse pair selected for Table II, with both atomic detunings fixed at zero. For each source, the Rabi frequency in the first time interval was fixed at zero, and QRT-GRAPE adjusted the remaining 99 values within , giving 198 adjustable values in total.
Following the general objective function in Eq. (37), we maximized the penalized objective function
where is the objective function defined in the main text and
measures variation between neighboring Rabi frequencies. Here, is the Rabi frequency for source in time interval , , and . The sum includes the change from the first fixed zero value to the next value. It excludes the change from the final driven interval to the additional decay interval, during which both Rabi frequencies are fixed at zero.
The corresponding gradient used by the optimizer is
The derivative of is evaluated using Eq. (41), and the derivative of is obtained by differentiating Eq. (64). We compared , , and .
Figure 7 shows how the smoothness penalty changes the optimized Rabi frequencies and the resulting objective function value .
Figure 7. Effect of penalizing variations between neighboring Rabi frequencies during optimization. All three optimizations started from the Gaussian Rabi pulse pair selected for Table II, with both atomic detunings fixed at zero. (a), (b) Optimized Rabi frequencies for sources A and B, respectively, for , , and . (c) Objective function value , evaluated with the penalty. (d) Measure of variation between neighboring Rabi frequencies, shown on a logarithmic scale. Shaded regions in (a) and (b) indicate the additional decay interval after the Rabi pulses are turned off. The change in each Rabi frequency when the pulses are turned off is excluded from .
Without the penalty, the optimized pulses gave and . For , decreased to 2.543, while remained at 0.2855. Increasing the coefficient to further reduced to 1.023, with . Relative to , the latter pulses reduced by 97.3%, with a 0.242% decrease in . All three calculations satisfied the L-BFGS-B convergence criterion.
Including the penalty thus allowed QRT-GRAPE to find smoother Rabi pulses while retaining nearly the same objective function value. The coefficient sets the balance between suppressing variations in the Rabi frequencies and maximizing , providing a way to account for pulse smoothness when designing controls for experimental implementation.
Sensitivity to Initial Detunings
We examined whether optimizing the atomic detunings together with the Rabi frequencies improved the final objective function value . All runs started from the Gaussian Rabi pulse pair selected for Table II, defined in Eq. (61) and listed in Table IV. We compared the final objective function values with , the existing result obtained by optimizing the Rabi frequencies alone from the same Gaussian pulse pair, with both atomic detunings fixed at zero.
We performed 30 optimizations in which both Rabi frequencies and both atomic detunings were adjusted. In each run, the initial atomic detuning values for sources A and B in the 99 adjustable time intervals were drawn independently from a uniform distribution over , without smoothing, rescaling, or clipping. All runs used the same initial Gaussian Rabi pulse pair. The value in the first time interval of each function was fixed at zero, and all four functions were set to zero for . Each run optimized 99 values of each Rabi frequency and atomic detuning, giving 396 adjustable values with bounds .
For run , we define the change from this existing result as
Figure 8 shows the differences for all 30 runs.
Figure 8. Comparison of objective function values from 30 optimizations of both Rabi frequencies and atomic detunings. All runs started from the same Gaussian Rabi pulse pair, with initial atomic detuning values drawn independently from a uniform distribution over in each adjustable time interval. The plotted difference is , where is the objective function value at termination and is the existing result obtained by optimizing the Rabi frequencies alone from the same Gaussian pulse pair, with both atomic detunings fixed at zero. The dashed line marks zero difference and is annotated with . Circles denote the 21 runs for which L-BFGS-B reported convergence; crosses denote the nine runs that reached the 500-iteration limit.
None of the 30 optimizations of both Rabi frequencies and atomic detunings exceeded the existing result obtained by optimizing the Rabi frequencies alone. The differences ranged from to , with a median of . These values include all 30 runs, including the nine that reached the iteration limit.
The existing solution obtained by optimizing the Rabi frequencies alone, with both atomic detunings fixed at zero, remains feasible when both the Rabi frequencies and atomic detunings are optimized. Adding atomic detunings as optimization variables therefore cannot lower the maximum attainable objective function value. In all 30 runs, we optimized both the Rabi frequencies and atomic detunings, starting from the initial Gaussian Rabi pulse pair and random initial atomic detuning values. The previously optimized Rabi pulses were not used as the initial values. These different initial conditions may have led the optimization along paths that did not reach the existing objective function value within the iteration limit and stopping criteria used here. Since optimizing both the Rabi frequencies and atomic detunings provided no improvement in this comparison, we fixed both atomic detunings at zero for simplicity in the numerical examples in the main text.
Numerical Verification of the Analytic Gradient
We evaluated the gradient for one fixed set of Rabi frequencies and atomic detunings. The physical parameters were those listed in Table 1. For the two sources, we collect the control values defined in the main text as
Here, and are the Rabi frequency and atomic detuning, respectively, for source in time interval . Each function is represented by 100 values during the driving period , with one value held constant over each interval of width . In the remaining 99 intervals, the Rabi frequencies were fixed at and , and both atomic detunings were fixed at 0.01. Both Rabi frequencies and both atomic detunings were fixed at zero during the additional decay interval, which was divided into 10 intervals.
For each selected Rabi frequency or atomic detuning value , we compared the analytic derivative with the centered finite-difference estimate
Here, denotes the vector of control values, and $\boldsymbol{e}_{m,k}^{(\mu)} denotes the standard basis vector whose component corresponding to is one and whose other components are all zero. Thus, the finite difference changes only that control value. We selected the time intervals
for each Rabi frequency and atomic detuning of sources A and B, giving 20 derivatives to compare in total. The values in the first interval () and the additional decay interval were fixed at zero and were not varied in the finite-difference calculation.
We quantified the difference between the analytic and finite-difference derivatives using the symmetric relative error
This measures the absolute difference relative to the larger magnitude of the two derivatives and is unchanged when the two derivative values are exchanged. The lower bound on the denominator prevents division by zero or by a very small value. We used 10 values of , equally spaced on a logarithmic scale within . Figure 9(a) shows the resulting errors for the 20 selected derivatives.
Figure 9. Numerical verification of the analytic QRT-GRAPE gradient. (a) Symmetric relative error between the analytic and centered finite-difference derivatives with respect to 20 selected Rabi frequency and atomic detuning values, as a function of the step size . Gray curves show all 20 selected derivatives, and colored curves highlight four of them. The values in square brackets in the legend are the time interval indices . (b) Median wall-clock time for evaluating and all 396 analytic derivatives, compared with that for computing all 396 derivatives by centered finite differences. Both methods were timed on the same computer with the same allocated CPU budget.
For each derivative, we first took the smallest error among the 10 tested step sizes:
where is the set of tested step sizes. The step size giving the smallest error differed among the derivatives. We then took the largest of these 20 minimum errors:
where the maximum is taken over the 20 selected derivatives. The agreement shown in Figure 9(a) confirms that our implementation correctly evaluates the analytic derivatives of with respect to the 20 tested Rabi frequency and atomic detuning values.
We compared the time required to evaluate and all 396 analytic derivatives with that required to compute the 396 derivatives by centered finite differences. Both calculations used the same fixed Rabi frequencies and atomic detunings, physical parameters, time grid, and trapezoidal integration. Centered finite differences required 792 evaluations of at one fixed step size . Both methods ran sequentially on Intel Xeon Silver 4410Y system running Linux under WSL2 in Docker, with the same allocated CPU budget of one logical CPU. For each method, we used one process, with one thread per process for matrix calculations. The software versions were Python 3.10.12, NumPy 1.26.4 [61], SciPy 1.15.1 [62], OpenBLAS 0.3.23.dev for NumPy, and OpenBLAS 0.3.28 for SciPy [63]. We first evaluated and its analytic gradient once and computed the full gradient by centered finite differences once. We then timed five further evaluations with each method and report the median time, excluding the initial evaluation. The timed operations were evaluation of and its analytic gradient for the analytic calculation, and all 792 objective function evaluations and assembly of the derivatives for centered finite differences.
Figure 9(b) compares the median times for the two calculations. The median time was 6.996 s for evaluating and its analytic gradient and 65.776 s for computing the gradient by centered finite differences. For this input, evaluating and its analytic gradient was 9.40 times as fast as computing the gradient by centered finite differences. This reduces the time spent evaluating gradients during pulse optimization.
References
References
- [1]K. E. Dorfman, F. Schlawin, and S. Mukamel, Nonlinear optical signals and spectroscopy with quantum light, Reviews of Modern Physics 88, 045008 (2016).DOI
- [2]R. Kubo, Statistical-Mechanical Theory of Irreversible Processes. I. General Theory and Simple Applications to Magnetic and Conduction Problems, Journal of the Physical Society of Japan 12, 570 (1957).
- [3]M. L. Baez, M. Goihl, J. Haferkamp, J. Bermejo-Vega, M. Gluza, and J. Eisert, Dynamical structure factors of dynamical quantum simulators, Proceedings of the National Academy of Sciences 117, 26123 (2020).DOI
- [4]A. T. Boothroyd, Principles of Neutron Scattering from Condensed Matter (Oxford University Press, 2020).DOI
- [5]K. Takase, M. Okada, T. Serikawa, S. Takeda, J.-i. Yoshikawa, and A. Furusawa, Complete temporal mode characterization of non-gaussian states by a dual homodyne measurement, Phys. Rev. A 99, 033832 (2019).
- [6]A. H. Kiilerich and K. Mølmer, Input-output theory with quantum pulses, Phys. Rev. Lett. 123, 123604 (2019).DOI
- [7]R. J. Glauber, The Quantum Theory of Optical Coherence, Physical Review 130, 2529 (1963).
- [8]C. Gardiner and P. Zoller, Quantum Noise: A Handbook of Markovian and Non-Markovian Quantum Stochastic Methods with Applications to Quantum Optics (Springer Science & Business Media, 2004).
- [9]D. Walls and G. J. Milburn, Quantum Coherence and Measurement Theory, in Quantum Optics, edited by D. Walls and G. J. Milburn (Springer, Berlin, Heidelberg, 2008).
- [10]H. J. Kimble, M. Dagenais, and L. Mandel, Photon Antibunching in Resonance Fluorescence, Physical Review Letters 39, 691 (1977).
- [11]K. A. Fischer, K. Müller, K. G. Lagoudakis, and J. Vučković, Dynamical modeling of pulsed two-photon interference, New Journal of Physics 18, 113053 (2016).
- [12]T. Walker, Improving the indistinguishability of single photons from an ion-cavity system, Physical Review A 102, 10.1103/PhysRevA.102.032616 (2020).DOI
- [13]S. Kikura, R. Asaoka, M. Koashi, and Y. Tokunaga, High-purity single-photon generation based on cavity qed, Phys. Rev. Res. 7, 013251 (2025).DOI
- [14]K. Tanji, H. Takahashi, W. Roga, and M. Takeoka, Rate-fidelity tradeoff in cavity-based remote entanglement generation, Physical Review A 110, 042405 (2024).DOI
- [15]M. H. Appel, A. Tiranov, S. Pabst, M. L. Chan, C. Starup, Y. Wang, L. Midolo, K. Tiurev, S. Scholz, A. D. Wieck, A. Ludwig, A. S. Sørensen, and P. Lodahl, Entangling a Hole Spin with a Time-Bin Photon: A Waveguide Approach for Quantum Dot Sources of Multiphoton Entanglement, Physical Review Letters 128, 233602 (2022).
- [16]T. K. Bracht, F. Kappe, M. Cygorek, T. Seidelmann, Y. Karli, V. Remesh, G. Weihs, V. M. Axt, and D. E. Reiter, Theory of time-bin-entangled photons from quantum emitters, Physical Review A 110, 067309 (2024).
- [17]L.-M. Duan, M. D. Lukin, J. I. Cirac, and P. Zoller, Long-distance quantum communication with atomic ensembles and linear optics, Nature 414, 413 (2001).DOI
- [18]N. Sangouard, C. Simon, H. de Riedmatten, and N. Gisin, Quantum repeaters based on atomic ensembles and linear optics, Reviews of Modern Physics 83, 33 (2011).
- [19]J. Hofmann, M. Krug, N. Ortegel, L. Gérard, M. Weber, W. Rosenfeld, and H. Weinfurter, Heralded Entanglement Between Widely Separated Atoms, Science 337, 72 (2012).DOI
- [20]H. Bernien, B. Hensen, W. Pfaff, G. Koolstra, M. S. Blok, L. Robledo, T. H. Taminiau, M. Markham, D. J. Twitchen, L. Childress, and R. Hanson, Heralded entanglement between solid-state qubits separated by three metres, Nature 497, 86 (2013).
- [21]V. Krutyansky, M. Galli, V. Krcmarsky, S. Baier, D. A. Fioretto, Y. Pu, A. Mazloom, P. Sekatski, M. Canteri, M. Teller, J. Schupp, J. Bate, M. Meraner, N. Sanguoard, B. P. Lanyon, and T. E. Northup, Entanglement of Trapped-Ion Qubits Separated by 230 Meters, Physical Review Letters 130, 050803 (2023).
- [22]T. van Leent, M. Bock, F. Fertig, R. Garthoff, S. Eppelt, Y. Zhou, P. Malik, M. Seubert, T. Bauer, W. Rosenfeld, W. Zhang, C. Becher, and H. Weinfurter, Entangling single atoms over 33 km telecom fibre, Nature 607, 69 (2022).DOI
- [23]K. Tanji, H. Shimizu, and M. Takeoka, High-rate qutrit entanglement swapping with photon-number basis, Phys. Rev. Res. 8, 013316 (2026).arxiv.org/abs/2507.04322
- [24]H. Aoki, K. Tanji, S. Kiryu, M. Takeoka, and J. Ishii-Hayase, Remote gottesman-kitaev-preskill entanglement generation between trapped ions (2026), arXiv:2610.21754 [quant-ph].arxiv.org/abs/2610.21754
- [25]N. H. Nickerson, Y. Li, and S. C. Benjamin, Topological quantum computing with a very noisy network and local error rates approaching one percent, Nature Communications 4, 1756 (2013).DOI
- [26]C. Monroe, R. Raussendorf, A. Ruthven, K. R. Brown, P. Maunz, L.-M. Duan, and J. Kim, Large-scale modular quantum-computer architecture with atomic memory and photonic interconnects, Physical Review A 89, 022317 (2014).DOI
- [27]N. H. Nickerson, J. F. Fitzsimons, and S. C. Benjamin, Freely Scalable Quantum Technologies Using Cells of 5-to-50 Qubits with Very Lossy and Noisy Photonic Links, Physical Review X 4, 041041 (2014).DOI
- [28]Y. Li and J. D. Thompson, High-Rate and High-Fidelity Modular Interconnects between Neutral Atom Quantum Processors, PRX Quantum 5, 020363 (2024).arxiv.org/abs/2401.04075
- [29]S. Sunami, S. Tamiya, R. Inoue, H. Yamasaki, and A. Goban, Scalable Networking of Neutral-Atom Qubits: Nanofiber-Based Approach for Multiprocessor Fault-Tolerant Quantum Computers, PRX Quantum 6, 010101 (2025).DOI
- [30]S. Wehner, D. Elkouss, and R. Hanson, Quantum internet: A vision for the road ahead, Science 362, eaam9288 (2018).
- [31]K. Azuma, S. E. Economou, D. Elkouss, P. Hilaire, L. Jiang, H.-K. Lo, and I. Tzitrin, Quantum repeaters: From quantum networks to the quantum internet, Reviews of Modern Physics 95, 045006 (2023).DOI
- [32]J. Sinclair, J. Ramette, B. Grinkemeyer, D. Bluvstein, M. D. Lukin, and V. Vuletić, Fault-tolerant optical interconnects for neutral-atom arrays, Physical Review Research 7, 013313 (2025).arxiv.org/abs/2408.08955
- [33]S. Singh, F. Gu, S. de Bone, E. Villaseñor, D. Elkouss, and J. Borregaard, Modular architectures and entangle ment schemes for error-corrected distributed quantum computation, npj Quantum Information 12, 3 (2025).
- [34]S. Singh, R. Kashiwagi, K. Tanji, W. Roga, D. Bhatti, M. Takeoka, and D. Elkouss, Fault-tolerant modular quantum computing with surface codes using single-shot emission-based hardware (2026), arXiv:2601.07241 [quant-ph].DOI
- [35]S. Kikura, K. Tanji, A. Goban, and S. Sunami, Passive quantum interconnects: Multiplexed remote entanglement generation with cavity-assisted photon scattering, Physical Review Applied 26, 034021 (2026).DOI
- [36]L.-Z. Liu, Y.-Z. Zhang, Z.-D. Li, R. Zhang, X.-F. Yin, Y.-Y. Fei, L. Li, N.-L. Liu, F. Xu, Y.-A. Chen, and J.-W. Pan, Distributed quantum phase estimation with entangled photons, Nature Photonics 15, 137 (2021).DOI
- [37]Y. Ueda, M. Ishihara, W. Roga, and M. Takeoka, Private quantum network sensing with efficient multipartite entanglement distribution via lossy channels, https://arxiv.org/abs/2505.10148 (2025).
- [38]S. J. Glaser, U. Boscain, T. Calarco, C. P. Koch, W. Kӧckenberger, R. Kosloff, I. Kuprov, B. Luy, S. Schirmer, T. Schulte-Herbrüggen, D. Sugny, and F. K. Wilhelm, Training Schrödinger’s cat: quantum optimal control, The European Physical Journal D 69, 279 (2015).DOI
- [39]C. P. Koch, U. Boscain, T. Calarco, G. Dirr, S. Filipp, S. J. Glaser, R. Kosloff, S. Montangero, T. Schulte-Herbrüggen, D. Sugny, and F. K. Wilhelm, Quantum optimal control in quantum technologies. Strategic report on current status, visions and goals for research in Europe, EPJ Quantum Technology 9, 19 (2022).
- [40]N. Khaneja, T. Reiss, C. Kehlet, T. Schulte-Herbrüggen, and S. J. Glaser, Optimal control of coupled spin dynamics: design of NMR pulse sequences by gradient ascent algorithms, Journal of Magnetic Resonance 172, 296 (2005).pubmed.ncbi.nlm.nih.gov/15649756
- [41]Y. Chen, Y. Hao, Z. Wu, B.-Y. Wang, R. Liu, Y. Hou, J. Cui, M.-H. Yung, and X. Peng, Iterative gradient ascent pulse engineering algorithm for quantum optimal control (2022), version: 2, 2212.02806 [quant-ph].arxiv.org/abs/2212.02806
- [42]Y. Lu, S. Joshi, V. San Dinh, and J. Koch, Optimal control of large quantum systems: assessing memory and runtime performance of GRAPE, Journal of Physics Communications 8, 025002 (2024).DOI
- [43]E. Khaleghian, A. F. Lipaei, A. Bahrampour, M. Nikaeen, and A. Bahrampour, Development of neural network-based optimal control pulse generator for quantum logic gates using the GRAPE algorithm in NMR quantum computer, Quantum Information Processing 25, 131 (2026).
- [44]D. Lewis, R. Wiersema, and S. Bose, Quantum Optimal Control with Geodesic Pulse Engineering (2025), arXiv:2508.16029 [quant-ph].arxiv.org/abs/2508.16029
- [45]R. Malarchick, When does numerical pulse optimization actually help? Error budgets, robustness tradeoffs, and calibration guidance for transmon single-qubit gates (2025), arXiv:2511.12799 [quant-ph].
- [46]J. Li, X. Yang, X. Peng, and C.-P. Sun, Hybrid Quantum-Classical Approach to Quantum Optimal Control, Physical Review Letters 118, 150503 (2017).DOI
- [47]D. Dong, Learning Control of Quantum Systems, in Encyclopedia of Systems and Control (Springer, London, 2020) pp. 1–7.
- [48]V. Petruhanov and A. Pechen, Incoherent GRAPE for Optimization of Quantum Systems with Environmentally Assisted Control, Russian Microelectronics 52, S424 (2023).DOI
- [49]V. N. Petruhanov and A. N. Pechen, GRAPE optimization for open quantum systems with time-dependent decoherence rates driven by coherent and incoherent controls, Journal of Physics A: Mathematical and Theoretical 56, 305303 (2023).DOI
- [50]Z.-J. Chen, H. Huang, L. Sun, Q.-X. Jie, J. Zhou, Z. Hua, Y. Xu, W. Wang, G.-C. Guo, C.-L. Zou, L. Sun, and X.-B. Zou, Robust and optimal control of open quantum systems, Science Advances 11, eadr0875 (2025).DOI
- [51]M. Lax, Formal Theory of Quantum Fluctuations from a Driven State, Physical Review 129, 2342 (1963).DOI
- [52]G. Guarnieri, A. Smirne, and B. Vacchini, Quantum regression theorem and non-Markovianity of quantum dynamics, Physical Review A 90, 022110 (2014), arXiv:1406.5340 [quant-ph].DOI
- [53]D. Lonigro and D. Chruściński, Quantum regression beyond the Born–Markov approximation for generalized spin-boson models, Physical Review A 105, 052435 (2022).DOI
- [54]C. Nölleke, A. Neuzner, A. Reiserer, C. Hahn, G. Rempe, and S. Ritter, Efficient teleportation between remote single-atom quantum memories, Physical Review Letters 110, 140403 (2013).pubmed.ncbi.nlm.nih.gov/25166964
- [55]Q. Ansel, E. Dionis, F. Arrouas, B. Peaudecerf, S. Guérin, D. Guéry-Odelin, and D. Sugny, Introduction to theoretical and experimental aspects of quantum optimal control, Journal of Physics B: Atomic, Molecular and Optical Physics 57, 133001 (2024).DOI
- [56]G. Lindblad, On the generators of quantum dynamical semigroups, Communications in Mathematical Physics 48, 119 (1976).DOI
- [57]V. Gorini, A. Kossakowski, and E. C. G. Sudarshan, Completely positive dynamical semigroups of N-level systems, Journal of Mathematical Physics 17, 821 (1976).DOI
- [58]W. Wirtinger, Zur formalen Theorie der Funktionen von mehr komplexen Veränderlichen, Mathematische Annalen 97, 357 (1927).DOI
- [59]K. Kreutz-Delgado, The Complex Gradient Operator and the CR-Calculus (2009), arXiv:0906.4835 [math.OC].
- [60]H. Goto, S. Mizukami, Y. Tokunaga, and T. Aoki, Figure of merit for single-photon generation based on cavity quantum electrodynamics, Physical Review A 99, 053843 (2019).DOI
- [61]C. R. Harris, K. J. Millman, S. J. van der Walt, et al., Array programming with NumPy, Nature 585, 357 (2020).
- [62]P. Virtanen, R. Gommers, T. E. Oliphant, et al., SciPy 1.0: Fundamental algorithms for scientific computing in python, Nature Methods 17, 261 (2020).
- [63]OpenBLAS, OpenBLAS.







