Systems and methods for digital-analog quantum machine learning

WO2024147803A3PCT designated stage expired Publication Date: 2024-11-07QUERA COMPUTING INC
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
PCT/US2023/031345
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2022-08-30
Filing Date
2023-08-29
Publication Date
2024-11-07

AI Technical Summary

Technical Problem

Current quantum machine learning methods face challenges in efficiently leveraging limited quantum resources to solve classically difficult problems, particularly in the noisy intermediate-scale era, where practical algorithms are needed to perform effectively with neutral atom arrays.

Method used

A hybrid digital-analog quantum machine learning approach using neutral atoms with Rydberg states, combining time-evolution and parameterized gates to optimize a loss function based on measurement results, allowing for both local and global entanglement operations.

Benefits of technology

This approach enables efficient learning of quantum and classical information, potentially outperforming classical methods by providing an exponential speedup in information processing and achieving high accuracy in tasks like phase detection and image classification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2023031345_07112024_PF_FP_ABST
    Figure US2023031345_07112024_PF_FP_ABST
Patent Text Reader

Abstract

Techniques for hybrid digital- analog quantum computation are described, which include systems and methods for controlling a quantum system to perform machine learning tasks. The quantum system can be trained in a manner comparable to classical machine learning classifiers by performing a series of parameterized operations (e.g., gates) on qubits of the quantum system, and subsequently performing measurements of at least some of the qubits. Results of the measurements may be used to calculate a loss function or other metric that can be optimized by varying at least some parameters of the operations. By determining the optimal values of the optimized parameters for a given task, the quantum system may be controlled using those values of the optimized parameters for various input test data, thereby producing predicted results for a wide range of previously unseen input data.
Need to check novelty before this filing date? Find Prior Art

Description

Attorney Docket No.: Q0075.70003WO00 SYSTEMS AND METHODS FOR DIGITAL-ANALOG QUANTUM MACHINE LEARNING BACKGROUND

[0001] Research in practical realizations of quantum computers has grown rapidly in recent years, due to their potential as quantum simulators and for outperforming their best-known classical counterparts on certain computational tasks. In the present noisy intermediate scale era, practical algorithms must solve classically difficult problems while leveraging very limited quantum resources. Among the most promising approaches to that end is the use of neutral atom arrays as the basis of the quantum computation. SUMMARY

[0002] According to some aspects, a method is provided of operating a quantum system comprising a plurality of neutral atoms exhibiting a first ground state, a Rydberg state, and first and second hyperfine ground states, the method comprising performing time-evolution of the plurality of neutral atoms with respect to the first ground state and the Rydberg state, applying a plurality of gates to the plurality of neutral atoms with respect to the first hyperfine ground state and the second hyperfine ground state, wherein the plurality of gates are parameterized by at least one parameter, measuring one or more of the plurality of neutral atoms, and calculating a value of a loss function based on one or more measurement results of said measurement of the one or more of the plurality of neutral atoms.

[0003] According to some aspects, a method is provided of training a quantum system comprising a plurality of neutral atoms exhibiting a first ground state, a Rydberg state, and first and second hyperfine ground states, the method comprising determining a plurality of input data parameters that represent training data, preparing the plurality of neutral atoms in a state parameterized by the determined plurality of input data parameters, iteratively controlling the plurality of neutral atoms and calculating a value of a loss function based on results of measuring one or more of the plurality of neutralAttorney Docket No.: Q0075.70003WO00 atoms to determine values of a plurality of training parameters that optimize the loss function, the iterative process comprising repeating steps of performing time-evolution of the plurality of neutral atoms with respect to the first ground state and the Rydberg state, applying a plurality of gates to the plurality of neutral atoms with respect to the first hyperfine ground state and the second hyperfine ground state, wherein the plurality of gates are parameterized by the plurality of training parameters, measuring one or more of the plurality of neutral atoms, and calculating the value of the loss function based on the results of measuring the one or more of the plurality of neutral atoms.

[0004] The foregoing apparatus and method embodiments may be implemented with any suitable combination of aspects, features, and acts described above or in further detail below. These and other aspects, embodiments, and features of the present teachings can be more fully understood from the following description in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0005] Various aspects and embodiments will be described with reference to the following figures. It should be appreciated that the figures are not necessarily drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures is represented by a like numeral. For purposes of clarity, not every component may be labeled in every drawing.

[0006] FIGs.1A-1B depict atomic states of^^^^, according to some embodiments;

[0007] FIGs.2A-2B depict ground and Rydberg states of^^^^, according to some embodiments;

[0008] FIGs.3A-3B depict illustrative phase diagrams for two different quantum systems each comprising an array of Rydberg atoms, according to some embodiments;

[0009] FIG.4 illustrates three phases shown in FIG.3A, according to some embodiments;

[0010] FIG.5 illustrates three phases shown in FIG.3B, according to some embodiments;Attorney Docket No.: Q0075.70003WO00

[0011] FIGs.6A-6B depict illustrative values of the detuning and the Rabi oscillation frequency during an adiabatic time evolution process, according to some embodiments;

[0012] FIG.7 depicts an example of adiabatic preparation in one dimension, according to some embodiments;

[0013] FIG.8 depicts an illustrative sequence of operations for a phase learning process for 4 qubits, according to some embodiments;

[0014] FIG.9 depicts an illustrative process of training a quantum neural network using a classical data set, according to some embodiments;

[0015] FIG.10 depicts an illustrative sequence of operations for classifying classical data, according to some embodiments;

[0016] FIG.11 shows results of a quantum machine learning process when trained independently on various one-shot locations in a single phase diagram of a 13 qubit chain, according to some embodiments;

[0017] FIG.12 shows results of a quantum machine learning process when trained independently on various one-shot locations in a single phase diagram of a 4x4 qubit lattice, according to some embodiments;

[0018] FIGs.13A-B depict results of a quantum machine learning process for two different quantum systems when using a random bitstring as a training target, according to some embodiments;

[0019] FIG.14 illustrates results of a quantum machine learning process for classifying classical image data, according to some embodiments;

[0020] FIG.15 illustrates an example of a computing system environment on which aspects of the invention may be implemented; and

[0021] FIG.16 is a schematic view of an illustrative apparatus on which aspects of the invention may be implemented. DETAILED DESCRIPTION

[0022] Techniques for hybrid digital-analog quantum computation are described. These techniques include systems and methods for controlling a quantum system to perform machine learning tasks, such as implementations in which the quantum systemAttorney Docket No.: Q0075.70003WO00 is used as a classifier (e.g., an architecture that approximates a desired function of input data). The quantum system can be trained in a manner comparable to classical machine learning classifiers by performing a series of parameterized operations (e.g., gates) on qubits of the quantum system, and subsequently performing measurements of at least some of the qubits. Results of the measurements may be used to calculate a loss function or other metric that can be optimized by varying at least some parameters of the operations. The optimization process may be performed with particular input training data, which may be represented by some of the parameters of the operations, an initial state of the system, or otherwise. By determining the optimal values of the optimized parameters for a given task, the quantum system may be controlled using those values of the optimized parameters for various input test data, thereby producing predicted results for a wide range of previously unseen input data.

[0023] According to some embodiments, acts of controlling the quantum system may comprise both analog and digital operations performed on qubits of the system. A “digital” operation, as referred to herein, may include the application of one or more discrete quantum gates each applied to one qubit, two qubits, or some other small number of qubits, whereas an “analog” operation, as referred to herein, may include time evolution (e.g., adiabatic time evolution) of all or a substantial portion (e.g., greater than 75%) of the quantum system according to physical laws.

[0024] According to some embodiments, the techniques described herein may be applied to a quantum system comprising neutral atom qubits, such as Rydberg atom qubits. The illustrative techniques described below are presented in relation to Rydberg atom qubits as an example, though it may be appreciated that these techniques are not necessarily limited to being used with Rydberg atom qubits, and may in at least some cases be applied to quantum systems comprising other types of qubits.

[0025] The formalism of neutral atom quantum computation is centered around Rydberg states. Consider an electrically neutral atom with a single electron of interest. In the ground state |^^, the electron is tightly bound to the nucleus. On the other hand, by providing a sufficient energy source, the electron may be excited to a large atomicAttorney Docket No.: Q0075.70003WO00 orbital, wherein it is so loosely bound that the electron is almost free. Such a state is referred to as a Rydberg state and is denoted |^^, as shown in FIGs.1A-1B.

[0026] FIG.1A depicts a two-photon process coupling the ground state of^^^^ atoms |^^ =|5^^ / ^, ^ = 2, ^^= 2^ to a Rydberg state |^^ =|70^^ / ^, ^ = 1 / 2, ^^= 1 / 2^ via an intermediate state |^^ =|6^^ / ^, ^ = 3, ^^= 3^ with detuning ^^using circularly polarized 420-nm and 1,013-nm lasers with single-photon Rabi frequencies !"^#and !^#^^respectively. FIG.1B depicts an effective two-level system$|^^ ↔|^^&, with Rabi frequency !'(= !"^#!^#^^ / 2^^and^).

[0027] Rydberg atoms are particularly promising as quantum computers because they have been shown to scale very well (e.g., into the hundreds and thousands of qubits) and because they can implement both digital and analog operations. With respect to digital operations, single-qubit rotation gates and various multi-qubit gates such as SWAP or controlled-NOT may be performed to manipulate one or more Rydberg atoms. With respect to analog operations, the one or more Rydberg atoms may be time-evolved according to the Schrödinger equation by controlling all (or most) of the atoms as an ensemble. Arrays of Rydberg atoms (e.g., 2D or 3D arrays) allow implementation of multi-qubit digital gates, providing one framework for local entanglement. Rydberg analog time evolution, however, offers a different and potentially advantageous form of entangling Rydberg atoms. Since time evolution can be implemented simply by applying lasers with parameters set to match those of a given Rydberg Hamiltonian on the atomic array, a quantum system can be operated to globally entangle the entire system (or a large fraction of the qubits of the system). Moreover, the time for which the lasers are applied can itself serve as a variational parameter, giving even more flexibility to the expressive power of the circuit.

[0028] The potential boosts provided by analog computation provide for quantum circuits on 1D, 2D or 3D arrays of Rydberg atoms (hereinafter “Rydberg arrays”) that combine both digital gates and time evolution. In the examples described herein, so- called hybrid digital-analog architectures are considered, in which control of the quantum system comprises intermixed (e.g., alternating) layers of digital single-qubit rotation gates and analog evolution. This approach may provide for both locally- andAttorney Docket No.: Q0075.70003WO00 globally-entangled operations using only efficient laser-based qubit manipulations on a scalable system of Rydberg atoms. This technique is in contrast with, for instance, a standard digital variational quantum circuit, in which the entanglement is achieved by some configuration of locally controlled multi-qubit gates, such as controlled-NOT, controlled-+, or controlled-SWAP. The techniques described herein instead use a hybrid digital-analog circuit to learn classical information, while a digital circuit is used for quantum information to learn input states prepared via analog computation.

[0029] As an implementation detail, when using Rydberg atoms as qubits, analog and digital operations may utilize different identifications of the qubit states |0^ and |1^ with different atomic states. In particular, the set of atomic ground states form a hyperfine ground state manifold. When digital gates are applied, qubit states |0^ and |1^ may be identified with two distinct states of the hyperfine ground state manifold, for which rotations and multi-qubit control gates can be more easily applied. These states are shown in FIG.2A. When performing analog operations, the qubit states |0^ and |1^ may instead be associated with a ground state |^^ and a Rydberg state |^^, as shown in FIG.1B, so that lasers may be operated with the Rydberg Hamiltonian parameters to implement state evolution over time. In some embodiments, a system may be configured to operate lasers to map between the hyperfine ground states |0^ and |1^ used for digital operations, and to operate lasers to map between the ground-Rydberg states |^^ and |^^ for analog operations.

[0030] FIG.2A depicts a secondary two-photon process coupling the two ground state manifolds of^^^^ namely|0^^ = 1, ^^= 0^ via an intermediate state|,^ =|5^^ / ^, ^ = 2, ^^= −1^ with a far off resonance^.using circularly polarized 795-nm lasers with single-photon Rabi frequencies !^01. FIG.2B depicts an effective three-level ladder system coupling= !"^#!^#^^ / 2^^and^), respectively.

[0031] In addition to efficient time evolution, Rydberg arrays also exhibit a physical phenomenon, known as the blockade effect, that is often useful for quantumAttorney Docket No.: Q0075.70003WO00 computation. This effect asserts that no two Rydberg atoms within a blockade radius ^3may both be in the Rydberg state |^^.

[0032] The blockade effect may provide for a natural approach to build multi-qubit quantum gates on Rydberg arrays. Since the blockade effect entangles atoms in the Rydberg state |^^, two atoms can be entangled during analog operations by exciting them to the Rydberg state and allowing them to interact by placing them within blockade proximity of one another. Similarly, two-qubit gates may be applied to Rydberg atoms as digital operations due to strong Rydberg interactions.

[0033] Techniques are described herein that apply a hybrid digital-analog approach to quantum machine learning. The analog power of Rydberg states may, for example, be utilized along with digital gates to learn both quantum and classical information.

[0034] As described further below, in a first illustrative application, digital Rydberg quantum systems may be controlled to learn the quantum phases of the Rydberg Hamiltonian, using input states produced through analog operation. The techniques described herein may have many applications, including but not limited to detection of exotic phases that may be of use in building quantum computers. Indeed, quantum computation can in theory give an exponential speedup to the learning of information encoded into quantum data, such as the amplitudes and phases of qubits.

[0035] In a second illustrative application, Rydberg digital-analog circuits may be operated to classify handwritten digits. Quantum computers may be able to outperform (or even substantially outperform) classical machine learning methods on classical (non- quantum) information, such as image data. The approach described herein provides a framework for operating a quantum system to perform classical machine learning, with an example described herein using Rydberg devices. Both of the illustrative applications allow for experimentation with presently available quantum technology.

[0036] Although the illustrative applications focus on two specific problems in quantum and classical information, respectively, the methods described to solve them may be highly general for quantum machine learning. For instance, the below-described anomaly detection technique may be applied to learn the functional landscape of classical and quantum information. In the case of the Rydberg system described herein,Attorney Docket No.: Q0075.70003WO00 the information may include the entanglement entropy of the ground state of a Rydberg Hamiltonian with respect to two parameters, but may also apply generally across many smooth functions with feature transitions. Similarly, the hybrid digital-analog circuit described below for learning of handwritten digits may be applied with appropriate modification to the hyperparameters and optimizer to classify encoding of general information—quantum or classical—into an input quantum state.

[0037] The above-described two illustrative applications are described further below. The first illustrative application is the learning of quantum phases, in which a mix of analog techniques for state preparation and digital techniques for quantum machine learning are applied to detect the quantum phase diagram of Rydberg states in an unsupervised manner. A first discussion below which relates to this illustrative application therefore describes a method of quantum machine learning on quantum information. The second discussion below relates to the second illustrative application in which a supervised quantum machine learning technique is applied to classify handwritten digits from the MNIST dataset (classical information) by use of a hybrid analog-digital circuit harnessing time evolution.

[0038] For purposes of illustration, the below description is focused on Rydberg quantum many-body phases in one and two dimensions. In one dimension, the Rydberg atoms form a chain of 4 qubits. In two dimensions, a general form is a geometric lattice array, though the case of a square lattice of 4 = 5^atoms may also be considered. In each case, the quantum structure of the Rydberg array may be characterized in a phase space of two parameters 6 and ^, which determine the Hamiltonian of the atom.

[0039] Specifically, the quantum many-body Hamiltonian on 4 Rydberg qubits is given by: Gwhere !)is the Rabi oscillation frequency on atom I, J)is a phase factor, ^)is the laser detuning frequency, DE)= |^)^^^)|, andAttorney Docket No.: Q0075.70003WO00 MLis the interaction potential, which follows a sixth-power law. The unit cell length 6 is implicitly encoded into F. In the subsequent description below, units are used in which ℏ = 1, and J)is set as J)= 0 so that the Rabi term operator is simply the Pauli spin operator P)Q. Moreover, all atoms are treated identically, so indexing on !, ^ is dropped. Given choices of ! and ^, the configuration of the Rydberg lattice will differ due to the Rydberg blockade effect, which states that no two atoms within the blockade radius ^3may be in the Rydberg state |^^, where F$^3&= !.

[0040] The Rydberg Hamiltonian may be analyzed over the phase space of dimensionless parameters ^ / ! and ^3 / 6. In particular, a phase diagram depicting the phase space of these two parameters may be constructed by obtaining the ground state of the Hamiltonian at each phase parameter, and then computing the entanglement entropy ^:where STUis the density matrix of the ground state of the Hamiltonian, bipartitioned into roughly equal-sized subsystems and STand SUare the reduced density matrices. Examples for 4 = 13 qubits in one dimension (with open boundary conditions) and 4 = 16 qubits arranged in a 4 × 4 lattice in two dimensions (with periodic boundary conditions) are shown in FIGs.3A and 3B, respectively, in which the entropies are computed by exact diagonalization (ED) of the respective Hamiltonians. The phase diagram in FIG.3A corresponds to a 13-qubit chain, whereas the phase diagram in FIG. 3B corresponds to a 4 x 4 lattice with periodic boundary conditions. In FIGs.3A-3B, the shading of a given point indicates the normalized entropy value, with darker spots indicating larger entropy values and lighter spots representing smaller entropy values.

[0041] One way to conceptualize the phase diagrams of Rydberg atoms as shown in FIGs.3A-3B is as follows. When the detuning frequency ^ is small, the most favored state is that in which most atoms are in the ground state, which leads to generallyAttorney Docket No.: Q0075.70003WO00 disordered phases. As ^ grows larger, more of the atoms may be excited to the |^^ state, but this behavior competes with the blockade effect, which limits the extent to which neighboring atoms can be in the |^^ state. As a result, as ^ increases, there are distinct favored phases corresponding to different ‘crystal’ configurations of the |^^ states.

[0042] In the two examples shown in FIGs.3A-3B, the lobes or regions of small ^, correspond to ordered phases, which are very close to a basis state in the computational basis. The remaining areas of the diagram are known as the disordered phase, and are separated from the ordered phases by a thin boundary of high-entanglement states. A useful method of visualization for these phases is the density number ];$ℋ& = |^gs$ℋ&|^;^|^= ^gs$ℋ&|DE;|gs$ℋ&^ where |gs$ℋ&^ is the ground state of ℋ and DE;= |^;^^^;| is the Rydberg projector on the _th qubit. ];$ℋ& being close to 0 or 1 indicates a classical configuration (that is, one near a bitstring basis state); otherwise, the qubit is highly entangled and therefore quantum.

[0043] In the one-dimensional chain, the ordered-phase lobes correspond to excitations of every D atoms along the chain (in the classical limit). Equivalently, the blockade radius is of D − 1 atoms. A visualization of the lowestphases is shown in FIG.4, with the darker squares corresponding to atoms in the Rydberg state ]$ℋ&≈ 1, and the lighter square corresponding to atoms in the ground state with ]$ℋ&≈ 0. Analogously, the two-dimensional square lattice also has ordered and disordered phases, which are known as the Neel, striated, and star phases. These are visualized, also via the density ]$ℋ&, in FIG.5.

[0044] The quantum machine learning techniques described herein allow for discovery of these phase diagrams with no a priori training or data. That is, in an unsupervised manner, the quantum circuit may be trained to distinguish ground states of distinct phases. These techniques thereby realize a quantum advantage in learning a widely applicable and important piece of quantum information.

[0045] Ground states of the Rydberg qubits may be prepared by a method of adiabatic evolution, where the quantum system is operated in an analog mode. The finite transformation corresponding to the standard Schrödinger equation is given byAttorney Docket No.: Q0075.70003WO00

[0046] The configuration wherein all atoms are in the state |^^ may be prepared and corresponds to a large negative detuning, e.g., ^∗= −2h × 10 MHz, since a negative detuning penalizes |^^ states due to the selected Hamiltonian ℋ. Thus, for a desired Rydberg lattice corresponding to a Hamiltonian with detuning ^, the state may be prepared by starting with |^i^ and then time-evolving the state adiabatically by adjusting ^$8& and !$8& appropriately. In some embodiments, for instance, a given time range may be divided into a large number (e.g., 300) of time slices jk, and a profile may be constructed for ^$8& in which ^ slowly increases from ^∗to the desired value. Similarly, the oscillation frequency ! may be gradually increased from 0 to its desired value, and then driven back to 0 near the end of the preparation process. At each time slice, the time evolution operator may be computed via Krylov subspace matrix exponentiation, and applied to |c$8&^. Illustrative values for ! and ^ over time during such processes are shown in FIGs.6A and 6B, respectively.

[0047] An example of the adiabatic preparation for the ℤ"phase in one dimension is depicted in FIG.7.

[0048] The illustrative process for quantum machine learning of a phase diagram may comprise at least the following acts. First, states of the 4 Rydberg qubits are prepared by analog operations (specifically, analog adiabatic evolution). Subsequently, lasers are applied to the system to transfer the Rydberg states of the qubits back to their hyperfine ground states. The states of the qubits are then input into a digital quantum circuit for learning. The digital quantum circuit comprises a layer of rotation gates, followed by nearest-neighbor controlled-+ gates and another layer of rotation gates. The rotation gates are each applied to individual qubits, whereas the nearest-neighbor controlled-Z gates are applied between nearest neighbor qubits. In the case of a lattice of qubits, the controlled-Z gates may be applied between any set of pairs of the qubits; in either case, there may be $4 − 1& controlled-Z gates applied.

[0049] To perform the quantum machine learning process, these digital operations may be applied to the Rydberg qubits in this order as described further below. In thisAttorney Docket No.: Q0075.70003WO00 example, the only parameters of variation are the Euler angles of rotation of the rotation gates. The digital quantum circuit may be considered a type of quantum neural network (QNN) due to its similarities to classical neural networks. By adjusting the rotation angles, a suitably chosen loss function may be optimized (e.g., minimized).

[0050] In practice, therefore, the quantum machine learning process may comprise: a) performing time-evolution of the Rydberg atoms as an analog operation (as described above, the analog operations may be performed with respect to a ground state and the Rydberg state of the atoms); b) applying the digital quantum circuit to the Rydberg atoms, which comprises rotation gates and controlled-Z gates as digital operations; c) measuring the Rydberg atoms (e.g., via a projective measurement); and d) calculating a value of a loss function using results of the measurements. Acts (a)-(d) may be repeated while varying the rotation angles of the rotation gates to optimize the value of the loss function calculated in act (d). An illustrative example of the phase learning process for 4 qubits is shown in FIG.8, wherein the parameters determined through the training process are l =mn^, n^, n^, n", n^o, n^o, n^o, n"op

[0051] According to some embodiments, the Rydberg atoms can be transported by optical tweezers, an example of which are described below in relation to FIG.16. As such, the geometry of the Rydberg atoms in different computational steps can be reconfigured. Therefore, between steps (a) and (b) in the quantum machine learning process above, the Rydberg atoms’ spatial locations can be changed. This reconfiguring can potentially enhance the entanglement generation speed and aid the quantum machine learning tasks.

[0052] According to some embodiments, the digital operation may be based on unsupervised intentional overfitting, which may also be referred to as anomaly detection. In this scheme, a simple, highly general loss function may be defined and the QNN may be trained as described above while using a classical updating algorithm to vary the rotation angle parameters until the loss function is minimized.

[0053] According to some embodiments, the loss function may be defined on the circuit q parameterized by l asAttorney Docket No.: Q0075.70003WO00 C,where tuis the Hamming weight, ^$;&is the bitstring on the _th measurement, and a total of s iterations of the phase learning process described above may be performed to obtain an average of the measurements. Minimizing this loss equivalently constructs a quantum circuit that decouples the qubits in x and sends them to |0^, a generic and phase-agnostic process. In some embodiments, a small subset of qubits may be selected randomly (e.g., 2√4 out of 4 qubits), thereby fixing a set x of measurement qubits that remains constant throughout the algorithm. Each measurement obtains a bitstring ^ of size |x|.

[0054] In some embodiments, training of the QNN may be performed using a single point in the phase diagram as a target, which may substantially overfit the QNN to that point. The point in the phase diagram that is used as a target for training is conveyed through the analog operation, since ℋ$8& depends on both ^ and !, so a choice of a point in the phase diagram is implied by the choice of ℋ$8&. According to some embodiments, multiple points across the phase diagram may be simultaneously trained on independent threads to see which point produces the best learned phase diagram. This approach is described further below.

[0055] According to some embodiments, the parameters l may be updated after calculation of the loss function value according to a classical algorithm, such as, but not limited to, the Simultaneous Perturbation Stochastic Approximation (SPSA) or the Nakanishi-Fujii-Todo (NFT) algorithms.

[0056] Once the rotation angle parameters have been selected to optimize the loss function, the value of the loss function on a mesh of points across the phase diagram may be computed. Assuming that the phase is the most “natural” characterization of the Hamiltonian ground states, the QNN should have loss nearly zero for states in the same phase as that of the training point, and varying degrees of losses for states of different phases. Thus, the QNN can, with no guidance whatsoever, approximately reconstruct the phase diagram.Attorney Docket No.: Q0075.70003WO00

[0057] One illustrative approach to computing the phase diagram based on the optimized rotation angle parameters is as follows. The phase diagram may be discretized into a DQ× Dz(e.g., 27×27) mesh. At each point$^ / !, ^3 / 6&in the phase diagram, the ground state of the corresponding Hamiltonian may be prepared via adiabatic evolution with ℋ$8&defined according to the selected values of ^ and !. The QNN may also be initialized with random rotation angles. The phase learning process used may be the same as that used to train the QNN, except with training parameters fixed at the values determined during training (e.g., for 4 qubits, the phase learning process shown in FIG. 8).

[0058] As a second illustrative application, a QNN may be trained to learn a dataset of handwritten digits using image data. In this example, the dataset is the MNIST dataset of handwritten digits. The QNN is tasked with classifying an input image with the correct number. To reduce the dimensionality of the image, the image data is transformed using principal component analysis (PCA) to generate 2D real numbers… , }a&that represent the image and that, once normalized, serve as the qubit coefficients. Since the principal components are sorted by their contribution to the total variance of the dataset, the entropy of the dataset is spread bywith the _th qubit. m{;pmay be normalized separately to be in the rangeR0, hV, forming a final dataset mn;p and m~;p. An illustrative process is shown in FIG.9, with the first two acts being performed on a classical computer (obtaining the image and performing PCA), and the remaining acts being performed using the quantum system.

[0059] To allow for natural time evolution, |0^ is identified with |^^ from the Rydberg hyperfine ground state manifold and |1^ is identified with the Rydberg state |^^. By executing single-qubit rotations appropriately, the following product state may be prepared to input into the quantum circuit:

[0060] The states in the area of the Rydberg Hamiltonian phase diagram may be prepared with little influence of the blockade effect, so that the qubits are free to beAttorney Docket No.: Q0075.70003WO00 either |0^ or |1^ without regard to spatial geometric configurations of the Rydberg lattice. This region may be the most feasible from an experimental point of view. In practice, this implies that ^3 / 6 < 1 and ^ / ! ≈ 0.75.

[0061] The variational circuit associated with the MNIST classifier is given by t alternating layers of single-qubit digital rotations and analog time evolution, with parameters l the Euler angles, and 8 the evolution time. The time 8 and the Hamiltonian are fixed, varying only the angles. An example using 4 = 4 qubits and with t = 3 is shown in FIG.10.

[0062] Below, preliminary results are presented from each of the above-described illustrative applications. The results are obtained by classical simulation of a quantum computation.

[0063] Below is described a proof of concept for both a one-dimensional chain of 13 qubits and a two-dimensional 4 × 4 square lattice. The former is simulated with open boundary conditions, allowing for excitations at the boundary compatible with ℤ^, ℤ^, and ℤ"phases. The latter is simulated with periodic boundary conditions, which allows for a rich phase diagram of the Neel, striated, and star phases without an intractably large system size. Moreover, exact diagonalization is used to obtain the states in place of adiabatic preparation for the proof of concept, as the focus will be on learning the true diagram regardless of what the true diagram is (which will depend slightly on whether ED (exact diagonalization) or adiabatic preparation is used).

[0064] For benchmarking purposes, the quantum machine learning output with ED input states is first compared with those of the ED phase diagram from FIG.4. Results of the quantum machine learning output when trained independently on various one-shot locations in the phase diagram are shown in FIG.11. Each of the different phase diagrams in FIG.11 represents results produced using the training point marked with a star in the respective phase diagram.

[0065] It may be observed that it is preferable to utilize a training point in the phase diagram that is arranged within one of the ordered phases, although each phase diagram has three identifiable lobes corresponding to ℤ^, ℤ^and ℤ".Attorney Docket No.: Q0075.70003WO00

[0066] These results are produced via the same process used to produce the 1D chain results. Beginning with ED input states, FIG.12 shows the results trained on various points, including the Neel, striated, and star phases as well as some disordered phases. Similar to the one-dimensional case, the quantum machine learning outputs an accurate phase diagram when trained on an ordered phase. Training on disordered phases results in a less accurate phase diagram, but still outlines many of the key features.

[0067] In some cases, the input to the quantum machine learning process may be a random classical bitstring, and no-multi qubit gates may be applied. Results of this approach are shown in FIGs.13A and 13B for the 13-qubit chain, and 4x4 lattice Rydberg arrays, respectively.

[0068] A convenient classical simulation of the above-described process to learn a dataset of handwritten digits using image data may be implemented in the Yao Circuits package from Julia. Since Yao has built-in automatic differentiation on quantum circuits over expectation values of observables, a cross-entropy loss may be used for a batch of size 4:where ^Gis a one-hot encoding of the true label and ^Gis the distribution of the circuit output on the computational basis, taken from the reduced density matrix.

[0069] For simplicity, only the numbers 1 and 9 are used, and only the binary classification problem is considered. This can be upgraded to the full classification problem by measuring more qubits. The evolution time is fixed at 8 = 0.5 ^s and t = 2 digital-analog layers are used. The system was then trained on 1000 random images over 60 iterations and tested on 200 images. Results of this process are shown in the top half of FIG.14. To examine the quantum advantage – that is, the advantage given by entanglement generated by analog evolution rather than independent optimization of a product state, which can be simulated efficiently on classical computers – these results may be compared with training on a single layer of rotation gates in the bottom half of the figure. In general, multiple runs suggest that product-state evolution can do no betterAttorney Docket No.: Q0075.70003WO00 than about 70% accuracy, while even a very short-depth circuit with analog evolution can achieve a 90% or higher accuracy.

[0070] The variational scheme may be generalized by additionally varying parameters of the analog time evolution. Most generically on a Rydberg lattice, the parameters varied may be $!, ^, 6, 8&, but for simplicity only the results for varying 6 are presented. Note that variation of 8 is equivalent to varying !, ^, and 1 / 6Ltogether identically. Moreover, only the parameters of the first analog gate are varied out of the 10 total gates in FIG.14. It may be observed that a high accuracy can be achieved with such variation, but it may be expected that more variation will allow the circuit to serve as a more powerful (expressive-wise) QNN for more general classification problems, including the classification of the full MNIST dataset of all 10 handwritten digits.

[0071] According to some embodiments, the techniques may generally be applied to any suitable quantum system comprising any suitable qubits, and having any suitable Hamiltonian, and are not limited to the illustrative Rydberg example described above. In many cases, the techniques may be readily scalable to a desired number of qubits. While this is particularly true in the case of neutral atom qubits, it may also be true for other qubit technologies to which these techniques may be applied.

[0072] An illustrative implementation of a computer system 1500 that may be used to control a quantum system as described herein is shown in FIG.15. The computer system 1500 may include one or more processors 1510 and one or more non-transitory computer-readable storage media or storage devices (e.g., memory 1520 and one or more non-volatile storage media 1530). The processor 1510 may control writing data to and reading data from the memory 1520 and the non-volatile storage device 1530 in any suitable manner, as the aspects of the invention described herein are not limited in this respect. To perform the functionality and / or techniques described herein, the processor 1510 may execute one or more instructions stored in one or more computer-readable storage media (e.g., the memory 1520, storage media, etc.), which may serve as non- transitory computer-readable storage media storing instructions for execution by the processor 1510.Attorney Docket No.: Q0075.70003WO00

[0073] In connection with techniques for controlling a quantum system to perform machine learning tasks as described herein, code used to, for example, control one or more energy sources (e.g., lasers, microwave emitters, and / or other sources of electromagnetic radiation) to perform analog and / or digital operations on one or more qubits, may be stored on one or more computer-readable storage media of computer system 1500. Processor 1510 may execute any such code to provide any techniques for controlling a quantum system to perform machine learning tasks as described herein. Any other software, programs or instructions described herein may also be stored and executed by computer system 1500. It will be appreciated that computer code may be applied to any aspects of methods and techniques described herein.

[0074] An illustrative apparatus 1600 that may be used to control a quantum system as described herein is shown in FIG.16. The example of FIG.16 is provided in the context of optical trapping of neutral atoms, although it will be appreciated that the techniques described herein are not necessarily limited to use with neutral atoms, nor with optical trapping of neutral atoms. As such, the below description of optical trapping of neutral atoms for quantum computation is provided merely as an illustrative example.

[0075] Optical trapping of neutral atoms is a powerful technique for isolating atoms in vacuum. Atoms are polarizable, and the oscillating electric field of a light beam induces an oscillating electric dipole moment in the atom. The associated energy shift in an atom from the induced dipole, averaged over a light oscillation period, is called the AC Stark shift. Based on the AC Stark shift induced by light that is detuned (i.e., offset in wavelength) from atomic resonance transitions, atoms are trapped at local intensity maxima (for red detuned, that is, longer wavelength trap light), because the atoms are attracted to light below the resonance frequency. The AC Stark shift is proportional to the intensity of the light. Thus, the shape of the intensity field is the shape of an associated atom trap. Optical tweezers utilize this principle by focusing a laser to a micron-scale waist, where individual atoms are trapped at the focus. Two-dimensional (2D) arrays of optical tweezers are generated by, for example, illuminating a spatial light modulator (SLM), which imprints a computer-generated hologram on the wavefront of the laser field. The 2D array of optical tweezers is overlapped with a cloud of laser- cooled atoms in a magneto-optical trap (MOT). The tightly focused optical tweezersAttorney Docket No.: Q0075.70003WO00 operate in a “collisional blockade” regime, in which single atoms are loaded from the MOT, while pairs of atoms are ejected due to light-assisted collisions, ensuring that the tweezers are loaded with at most single atoms, but the loading is probabilistic, such that the trap is loaded with a single atom with a probability of about 50-60%.

[0076] To prepare deterministic atom arrays, a real-time feedback procedure identifies the randomly loaded atoms and rearranges them into pre-programmed geometries. Atom rearrangement requires moving atoms in tweezers which can be smoothly steered to minimize heating, by using, for example, acousto-optic deflectors (AODs) to deflect a laser beam by a tunable angle which is controlled by the frequency of an acoustic waveform applied to the AOD crystal. Dynamic tuning of the acoustic frequency translates into smooth motion of an optical tweezer. A multi-frequency acoustic wave creates an array of laser deflections, which, after focusing through a microscope objective, forms an array of optical tweezers with tunable position and amplitude that are both controlled by the acoustic waveform. Atoms are rearranged by using an additional set of dynamically moving tweezers that are overlaid on top of the SLM tweezer array.

[0077] Optical tweezer arrays are a powerful and flexible way to construct large scale systems composed of individual particles. Each optical tweezer traps a single particle, including, but not limited to, individual neutral atoms and molecules for applications in quantum technology. Loading individual particles into such tweezer arrays is a stochastic process, where each tweezer in the system is filled with a single particle with a finite probability p<1, for example p~0.5 in the case of many neutral atom tweezer implementations. To compensate for this random loading, real-time feedback may be obtained by measuring which tweezers are loaded and then sorting the loaded particles into a programmable geometry. This may be performed by moving one particle at a time, or in parallel.

[0078] Parallel sorting may be achieved by using two acousto-optic deflectors (AODs) to generate multiple tweezers that can pick up particles from an existing particle-trapping structure, move them simultaneously, and release them somewhere else. This can include moving particles around within a single trapping structure (e.g., tweezerAttorney Docket No.: Q0075.70003WO00 array) or transporting and sorting particles from one trapping system to another (e.g., between one tweezer array and another type of optical / magnetic trap). This sorting is flexible and allows programmed positioning of each particle. Each movable trap is formed by the AODs and its position is dynamically controlled by the frequency components of the radiofrequency (RF) drive field for the AODs. Since the RF drive of the AODs can be controlled in real time and can include any combination of frequency components, it is possible to generate any grid of traps (such as a line of arbitrarily positioned traps), move the rows or columns of the grid, and add or remove rows and columns of the grid, by changing the number, magnitude, and distribution of the frequency components in the RF drive fields of the AODs.

[0079] In an exemplary embodiment, an optical tweezer array is created using a liquid crystal on silicon spatial light modulator (SLM), which can programmatically create flexible arrangements of tweezers. These tweezers are fixed in space for a given experimental sequence and loaded stochastically with individual atoms, such that each tweezer is loaded with probability p ~ 0.5. A fluorescence image of the loaded atoms is taken, to identify in real-time which tweezers are loaded and which are empty.

[0080] After detecting which tweezers are loaded, movable tweezers overlapping the optical tweezer array can dynamically reposition atoms from their starting locations to fill a target arrangement of traps with near-unity filling. The movable tweezers are created with a pair of crossed AODs. These AODs can be used to create a single moveable trap which moves one atom at a time to fill the target arrangement or to move many atoms in parallel.

[0081] Referring to FIG.16, a schematic view is provided of an apparatus 1600 for quantum computation according to embodiments of the present disclosure. As shown in FIG.16, using a beam generated by a light source 1602 (for example, a coherent light source, in some example embodiments – a monochromatic light source), SLM 1604 forms an array of trapping beams (a tweezer array) which is imaged onto trapping plane 1608 in vacuum chamber 1610 by an optical train that, in the example embodiment shown in FIG.16, comprises elements 1606a, 1606c, 1606d, and a high numerical aperture (NA) objective 1606e. Other suitable optical trains can be employed. Using aAttorney Docket No.: Q0075.70003WO00 beam generated by light source 1612 (for example, a coherent light source; in some example embodiments - a monochromatic light source), a pair of AODs 1614 and 1616, having non-parallel directions of acoustic wave propagation (for example, orthogonal directions) creates dynamically movable sorting beams. By using the optical train, such as the one depicted in FIG.16 (elements 1617, 1606b, 1606c, 1606d, and 1606e), the sorting beams are overlapped with the trapping beams. It is understood that other optical train can be used to achieve the same result. For example, source 1602 and 1612 can be a single source, and the trapping beam and the sorting beam are generated by a beam splitter.

[0082] The dynamic movement of the steering beams is accomplished by employing two non-parallel AODs 1614, 1616, arranged in series. In the example embodiment depicted in FIG.16, one AOD defines the direction of “rows” (“horizontal” – the ‘X’ AOD) and the other AOD defines the direction of “columns” (“vertical” – the ‘Y’ AOD). Each AOD is driven with an arbitrary RF waveform from an arbitrary waveform generator 1620, which is generated in real-time by a computer 1622 (e.g., an instance of computer system 1500) which processes the feedback routine after analyzing the image of where atoms are loaded. If each AOD is driven with a single frequency component, then a single steering beam (“AOD trap”) is created in the same plane 1608 as the SLM trap array. The frequency of the X AOD drive determines the horizontal position of the AOD trap, and the frequency of the Y AOD drive determines the vertical position; in this way, a single AOD trap can be steered to overlap with any SLM trap.

[0083] In FIG.16, laser 1602 projects a beam of light onto SLM 1604. SLM 1604 can be controlled by computer 1622 in order to generate a pattern of beams (“trapping beams” or “tweezer array”). The pattern of beams is focused by lens 1606a, passes through mirror 1606b, and is collimated by lens 1606c on mirror 1606d. The reflected light passes through objective 1606e to focus an optical tweezer array in vacuum chamber 1610 on trapping plane 1608. The laser light of the optical tweezer array continues through objective 1624a, and passes through dichroic mirror 1624b to be detected by charge-coupled device (CCD) camera 1624c.Attorney Docket No.: Q0075.70003WO00

[0084] Vacuum chamber 1610 may be illuminated by an additional light source (not pictured). Fluorescence from atoms trapped on the trapping plane also passes through objective 1624a, but is reflected by dichroic mirror 1624b to electron- multiplying CCD (EMCCD) camera 1624d.

[0085] In this example, laser 1612 directs a beam of light to AODs 1614, 1616. AODs 1614, 1616 are driven by arbitrary wave generator (AWG) 1620, which is in turn controlled by computer 1622. Crossed AODs 1614, 1616 emit one or more beams as set forth above, which are directed to focusing lens 1617. The beams then enter the same optical train 1606b…1606e as described above with regard to the optical tweezer array, focusing on trapping plane 1608.

[0086] In the context of the system 1600 of FIG.16, the quantum states of neutral atoms held in the optical tweezer array produced by the SLM 1604 may be manipulated (e.g., initialized, modified through one or more gates, etc.) through the application of optical pulses to the neutral atoms. As one example, one or more optical pulses may be applied, by the light source 1602 or otherwise, the pulses having time durations and a frequency (or frequencies) selected so as to produce a desired change in the quantum state of one or more neutral atoms (e.g., to perform a desired single-qubit rotation gate of a neutral atom).

[0087] Various inventive concepts may be embodied as at least one non-transitory computer readable storage medium (e.g., a computer memory, one or more floppy discs, compact discs, optical discs, magnetic tapes, flash memories, circuit configurations in Field Programmable Gate Arrays or other semiconductor devices, etc.) or a computer readable storage device encoded with one or more programs that, when executed on one or more computers or other processors, implement some of the various embodiments of the present invention. The non-transitory computer-readable medium or media may be transportable, such that the program or programs stored thereon may be loaded onto any computer resource to implement various aspects of the present invention as discussed above.

[0088] The terms “program,” “software,” and / or “application” are used herein in a generic sense to refer to any type of computer code or set of computer-executableAttorney Docket No.: Q0075.70003WO00 instructions that can be employed to program a computer or other processor to implement various aspects of embodiments as discussed above. Additionally, it should be appreciated that according to one aspect, one or more computer programs that when executed perform methods of one or more embodiments described herein need not reside on a single computer or processor, but may be distributed in a modular fashion among different computers or processors to implement various aspects of the present invention.

[0089] Having thus described several aspects of at least one embodiment of this invention, it is to be appreciated that various alterations, modifications, and improvements will readily occur to those skilled in the art.

[0090] Such alterations, modifications, and improvements are intended to be part of this disclosure, and are intended to be within the spirit and scope of the invention. Further, though advantages of the present invention are indicated, it should be appreciated that not every embodiment of the technology described herein will include every described advantage. Some embodiments may not implement any features described as advantageous herein and in some instances one or more of the described features may be implemented to achieve further embodiments. Accordingly, the foregoing description and drawings are by way of example only.

[0091] The above-described embodiments of the techniques for controlling a quantum system to perform machine learning tasks described herein can be implemented in any of numerous ways. For example, the embodiments may be implemented using hardware, software or a combination thereof. When implemented in software, the software code can be executed on any suitable processor or collection of processors, whether provided in a single computer or distributed among multiple computers. Such processors may be implemented as integrated circuits, with one or more processors in an integrated circuit component, including commercially available integrated circuit components known in the art by names such as CPU chips, GPU chips, microprocessor, microcontroller, or co-processor. Alternatively, a processor may be implemented in custom circuitry, such as an ASIC, or semi-custom circuitry resulting from configuring a programmable logic device. As yet a further alternative, a processor may be a portion of a larger circuit or semiconductor device, whether commercially available, semi-customAttorney Docket No.: Q0075.70003WO00 or custom. As a specific example, some commercially available microprocessors have multiple cores such that one or a subset of those cores may constitute a processor. Though, a processor may be implemented using circuitry in any suitable format.

[0092] Various aspects of the present invention may be used alone, in combination, or in a variety of arrangements not specifically described in the embodiments described in the foregoing and is therefore not limited in its application to the details and arrangement of components set forth in the foregoing description or illustrated in the drawings. For example, aspects described in one embodiment may be combined in any manner with aspects described in other embodiments.

[0093] Also, the invention may be embodied as a method, of which an example has been provided. The acts performed as part of the method may be ordered in any suitable way. Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which may include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.

[0094] Use of ordinal terms such as “first,” “second,” “third,” etc., in the claims to modify a claim element does not by itself connote any priority, precedence, or order of one claim element over another or the temporal order in which acts of a method are performed, but are used merely as labels to distinguish one claim element having a certain name from another element having a same name (but for use of the ordinal term) to distinguish the claim elements.

[0095] The terms “approximately” and “about” may be used to mean within ±20% of a target value in some embodiments, within ±10% of a target value in some embodiments, within ±5% of a target value in some embodiments, and yet within ±2% of a target value in some embodiments. The terms “approximately” and “about” may include the target value. The term “substantially equal” may be used to refer to values that are within ±20% of one another in some embodiments, within ±10% of one another in some embodiments, within ±5% of one another in some embodiments, and yet within ±2% of one another in some embodiments.

[0096] The term “substantially” may be used to refer to values that are within ±20% of a comparative measure in some embodiments, within ±10% in someAttorney Docket No.: Q0075.70003WO00 embodiments, within ±5% in some embodiments, and yet within ±2% in some embodiments. For example, a first direction that is “substantially” perpendicular to a second direction may refer to a first direction that is within ±20% of making a 90° angle with the second direction in some embodiments, within ±10% of making a 90° angle with the second direction in some embodiments, within ±5% of making a 90° angle with the second direction in some embodiments, and yet within ±2% of making a 90° angle with the second direction in some embodiments.

[0097] Also, the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of “including,” “comprising,” or “having,” “containing,” “involving,” and variations thereof herein, is meant to encompass the items listed thereafter and equivalents thereof as well as additional items.

Claims

Attorney Docket No.: Q0075.70003WO00 CLAIMS What is claimed is:

1. A method of operating a quantum system comprising a plurality of neutral atoms exhibiting a first ground state, a Rydberg state, and first and second hyperfine ground states, the method comprising: performing time-evolution of the plurality of neutral atoms with respect to the first ground state and the Rydberg state; applying a plurality of gates to the plurality of neutral atoms with respect to the first hyperfine ground state and the second hyperfine ground state, wherein the plurality of gates are parameterized by at least one parameter; measuring one or more of the plurality of neutral atoms; and calculating a value of a loss function based on one or more measurement results of said measurement of the one or more of the plurality of neutral atoms.

2. The method of claim 1, wherein performing time-evolution of the plurality of neutral atoms comprises adiabatically time-evolving the plurality of neutral atoms.

3. The method of claim 1, wherein performing time-evolution of the plurality of neutral atoms comprises time-evolving the plurality of neutral atoms according to a preparation Hamiltonian.

4. The method of claim 1, wherein performing time-evolution of the plurality of neutral atoms comprises time-evolving the plurality of neutral atoms using the first ground state and the Rydberg state as computational basis states.

5. The method of claim 1, wherein applying the plurality of gates to the plurality of neutral atoms comprises applying the plurality of gates to the plurality of neutral atoms using the first hyperfine ground state and the second hyperfine ground state as computational basis states.Attorney Docket No.: Q0075.70003WO00 6. The method of claim 1, further including moving a selected plurality of the neutral atoms from a first spatial configuration into a second spatial configuration between performing time-evolution of the plurality of neutral atoms and applying the plurality of gates to the plurality of neutral atoms.

7. The method of claim 1, wherein measuring the one or more of the plurality of neutral atoms comprises applying one or more projective measurements on the one or more of the plurality of neutral atoms.

8. The method of claim 1, wherein the plurality of gates comprises a plurality of single-qubit gates.

9. The method of claim 1, wherein the plurality of gates comprises a plurality of multi-qubit gates.

10. The method of claim 9, wherein the multi-qubit gates comprise one or more two- qubit gates.

11. The method of claim 10, wherein the one or more two-qubit gates comprise one or more controlled-Z gates.

12. The method of claim 1, further comprising determining values of the at least one parameter that optimize the loss function.

13. The method of claim 12, wherein determining the values of the at least one parameter that optimize the loss function comprises repeating: performing time-evolution of the plurality of neutral atoms and applying the plurality of gates to the plurality of neutral atoms.

14. The method of claim 12, further comprising, for each of a plurality of different ground states of the quantum system each represented by one or more parameters:Attorney Docket No.: Q0075.70003WO00 performing time-evolution of the plurality of neutral atoms with respect to the first ground state and the Rydberg state; applying the plurality of gates to the plurality of neutral atoms with respect to the first hyperfine ground state and the second hyperfine ground state, wherein the plurality of gates are parameterized by the determined values of the at least one parameter that optimize the loss function; and measuring the plurality of neutral atoms.

15. The method of claim 14, further comprising generating a phase diagram based at least in part on measurement results of measuring the plurality of neutral atoms for each of the different ground states of the quantum system.

16. The method of claim 1, wherein the at least one parameter includes one or more rotation angle parameters.

17. The method of claim 1, wherein the at least one parameter includes one or more parameters of a computational Hamiltonian each selected from a Rabi frequency, a detuning, and a lattice constant.

18. The method of claim 1, wherein the at least one parameter includes one or more lengths of time for which to perform time-evolution of the plurality of neutral atoms according to a computational Hamiltonian.

19. The method of claim 1, comprising selecting values of one or more Hamiltonian parameters and comprising performing said time-evolution of the plurality of neutral atoms according to a preparation Hamiltonian of the quantum system given the selected values of the one or more Hamiltonian parameters.

20. The method of claim 19, wherein the one or more Hamiltonian parameters include a detuning frequency and a Rabi frequency.Attorney Docket No.: Q0075.70003WO00 21. The method of claim 19, wherein the plurality of neutral atoms comprise a plurality of Rydberg atoms on a lattice having a lattice constant, and wherein the one or more Hamiltonian parameters include the lattice constant.

22. The method of claim 1, wherein the loss function is a function of a Hamming weight of the one or more measurement results of said measurement of the one or more of the plurality of neutral atoms.

23. The method of claim 1, wherein the plurality of gates comprise a plurality of rotation operator gates, and wherein the at least one parameter that parameterizes the rotation operator gates includes at least one rotation angle for each of the rotation operator gates.

24. The method of claim 1, wherein applying the plurality of gates to the plurality of neutral atoms comprises applying a plurality of single-qubit gates to the plurality of neutral atoms.

25. The method of claim 24, wherein applying the plurality of gates to the plurality of neutral atoms further comprises, subsequent to applying the plurality of single-qubit gates to the plurality of neutral atoms, applying a plurality of two-qubit gates to pairs of the plurality of neutral atoms.

26. The method of claim 24, wherein the plurality of single-qubit gates include a first plurality of rotation operator gates and a second plurality of rotation operator gates, wherein the method comprises: applying the first plurality of rotation operator gates to the plurality of neutral atoms; subsequent to applying the first plurality of rotation operator gates to the plurality of neutral atoms, applying a plurality of multi-qubit gates to the plurality of neutral atoms; andAttorney Docket No.: Q0075.70003WO00 subsequent to applying the plurality of multi-qubit gates, applying a second plurality of rotation operator gates to the plurality of neutral atoms, and wherein the at least one parameter that parameterizes the plurality of single-qubit gates includes a rotation angle for each of the first plurality of rotation operator gates and a rotation angle for each of the second plurality of rotation operator gates.

27. The method of claim 26, wherein the plurality of multi-qubit gates are controlled- Z gates.

28. The method of claim 1, wherein applying the plurality of gates to the plurality of neutral atoms comprises: applying a first plurality of single-qubit gates to the plurality of neutral atoms; and subsequent to applying the plurality of single-qubit gates to the plurality of neutral atoms, applying a second plurality of single qubit gates to the plurality of neutral atoms.

29. The method of claim 1, wherein performing time-evolution of the plurality of neutral atoms with respect to the first ground state and the Rydberg state comprises performing time-evolution of the plurality of neutral atoms according to a preparation Hamiltonian.

30. The method of claim 29, wherein the preparation Hamiltonian is a time-varying waveform Hamiltonian that is a function of a detuning frequency and a Rabi frequency.

31. The method of claim 1, wherein the plurality of neutral atoms comprises a plurality of Rydberg atoms.

32. The method of claim 31, wherein the plurality of Rydberg atoms comprises a one-dimensional array of Rydberg atoms.

33. The method of claim 31, wherein the plurality of Rydberg atoms comprise a two- dimensional lattice of Rydberg atoms.Attorney Docket No.: Q0075.70003WO00 34. The method of claim 1, wherein measuring the one or more of the plurality of neutral atoms comprises randomly selecting a subset of the plurality of neutral atoms, and measuring the neutral atoms in the subset of the plurality of neutral atoms.

35. The method of claim 1, wherein applying the plurality of gates to the plurality of neutral atoms comprises directing one or more laser pulses onto each of the plurality of neutral atoms.

36. The method of claim 1, wherein the neutral atoms comprise an array of single neutral atoms in optical or magnetic traps.

37. A method of training a quantum system comprising a plurality of neutral atoms exhibiting a first ground state, a Rydberg state, and first and second hyperfine ground states, the method comprising: determining a plurality of input data parameters that represent training data; preparing the plurality of neutral atoms in a state parameterized by the determined plurality of input data parameters; iteratively controlling the plurality of neutral atoms and calculating a value of a loss function based on results of measuring one or more of the plurality of neutral atoms to determine values of a plurality of training parameters that optimize the loss function, the iterative process comprising repeating steps of: performing time-evolution of the plurality of neutral atoms with respect to the first ground state and the Rydberg state; applying a plurality of gates to the plurality of neutral atoms with respect to the first hyperfine ground state and the second hyperfine ground state, wherein the plurality of gates are parameterized by the plurality of training parameters; measuring one or more of the plurality of neutral atoms; and calculating the value of the loss function based on the results of measuring the one or more of the plurality of neutral atoms.Attorney Docket No.: Q0075.70003WO00 38. The method of claim 37, wherein the training data comprises image data, and where determining the plurality of input data parameters comprises determining a plurality of values that represents the image data.

39. The method of claim 38, wherein determining the plurality of values that represent the image data includes principal component analysis (PCA).

40. The method of claim 37, wherein the plurality of training parameters comprise a first plurality of training parameters and a second plurality of training parameters, and wherein the iterative process comprises repeating steps of: performing time-evolution of the plurality of neutral atoms for a first time period; applying a first plurality of single-qubit gates to the plurality of neutral atoms, wherein the first plurality of single-qubit gates are parameterized by a first plurality of training parameters; performing time-evolution of the plurality of neutral atoms for a second time period; and applying a second plurality of single-qubit gates to the plurality of neutral atoms, wherein the second plurality of single-qubit gates are parameterized by a second plurality of training parameters.

41. The method of claim 37, wherein the plurality of training parameters include one or more parameters selected from a Rabi frequency, a detuning and a lattice constant.

42. The method of claim 37, wherein the plurality of training parameters include one or more lengths of time for which to perform time-evolution of the plurality of neutral atoms according to a computational Hamiltonian.

43. The method of claim 37, wherein the state parameterized by the determined plurality of input data parameters is a product state.Attorney Docket No.: Q0075.70003WO00 44. The method of claim 43, wherein the product state is a superposition of the first ground state and the Rydberg state.

45. The method of claim 43, wherein the plurality of neutral atoms comprises a plurality of Rydberg atoms and wherein each qubit in the product state is in a superposition of the first hyperfine ground state and the second hyperfine ground state.

46. The method of claim 37, comprising calculating the loss function from a set of training labels and the results of measuring the one or more of the plurality of neutral atoms.

47. The method of claim 46, wherein the loss function is a cross-entropy between a distribution determined by the results of measuring the one or more of the plurality of neutral atoms and a one-hot encoding of the training labels.

48. The method of claim 37, wherein the plurality of gates comprise a plurality of rotation operator gates, and wherein the plurality of training parameters that parameterizes the rotation operator gates includes at least one rotation angle for each of the rotation operator gates.

49. The method of claim 48, wherein the rotation operator gates are 3-dimensional rotation operator gates, and wherein the plurality of training parameters that parameterizes the 3-dimensional rotation operator gates includes one or more Euler angles for each 3-dimensional rotation operator gate.

50. The method of claim 37, wherein performing time-evolution of the plurality of neutral atoms comprises adiabatically time-evolving the plurality of neutral atoms.

51. The method of claim 37, wherein performing time-evolution of the plurality of neutral atoms comprises time-evolving the plurality of neutral atoms according to a preparation Hamiltonian.Attorney Docket No.: Q0075.70003WO00 52. The method of claim 37, wherein performing time-evolution of the plurality of neutral atoms comprises time-evolving the plurality of neutral atoms using the first ground state and the Rydberg state as computational basis states.

53. The method of claim 37, wherein applying the plurality of gates to the plurality of neutral atoms comprises applying the plurality of gates to the plurality of neutral atoms using the first hyperfine ground state and the second hyperfine ground state as computational basis states.

54. The method of claim 37, wherein measuring the one or more of the plurality of neutral atoms comprises applying one or more projective measurements on the one or more of the plurality of neutral atoms.

55. The method of claim 37, wherein the plurality of gates comprises a plurality of single-qubit gates.

56. The method of claim 37, wherein the plurality of gates comprises a plurality of multi-qubit gates.

57. The method of claim 56, wherein the multi-qubit gates comprise one or more two-qubit gates.

58. The method of claim 57, wherein the one or more two-qubit gates comprise one or more controlled-Z gates.

59. The method of any preceding claim, wherein the first ground state is the same as the second hyperfine ground state.

Citation Information

Patent Citations

  • Solving a set of (NON)linear differential equations using a hybrid data processing system comprising a classical computer system and a quantum computer system

    WO2022101483A1