Method and system for measuring similarity between two graphs

By employing an atomic array to model graphs and calculating feature vectors through quantum processing, the method addresses the limitations of existing quantum hardware and graph kernels, enabling accurate similarity measurement and classification for graphs with varying node counts.

JP7580010B2Active Publication Date: 2024-11-08PASQAL SAS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
JP2024501152
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2021-07-07
Filing Date
2022-07-07
Publication Date
2024-11-08
Estimated Expiration
2042-07-07

AI Technical Summary

Technical Problem

Existing quantum hardware for measuring graph similarity is limited to graphs with a predetermined number of nodes and struggles with graphs having a large number of nodes, and existing graph kernels lack optimal accuracy.

Method used

A method using an atomic array to model graphs, where each atom represents a node and interaction probability represents an edge, with a repeatable measurement process to calculate feature vectors, allowing for graphs with varying numbers of nodes and improved accuracy through quantum processing.

Benefits of technology

The method provides feature vectors with enhanced expressivity, enabling accurate similarity measurement and classification tasks, particularly for graphs with over 10 nodes, achieving higher F-scores compared to classical methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007580010000001
    Figure 0007580010000001
  • Figure 0007580010000002
    Figure 0007580010000002
  • Figure 0007580010000003
    Figure 0007580010000003
Patent Text Reader

Abstract

A method is provided for measuring the similarity between two graphs even when the graphs have a large number of nodes. The method includes the steps of: determining a feature vector for each of two graphs; and calculating an estimate of the similarity between the two graphs by comparing the two feature vectors. In determining the feature vector for each graph, a process including the steps of: providing an atomic array of atoms having a plurality of energy levels, each of which has at least a first reference state and a first excited state, arranged to model the graph, preparing the atoms to be in a reference state, applying an interaction sequence to the atomic array, detecting atoms in an excited state, and calculating observables of the atomic array is repeated a number of times to obtain a number of observables, and a feature vector is calculated based on the numerical distribution of the observables.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present specification relates to a method and system for measuring the similarity between two graphs. [Background technology]

[0002] Many systems that contain relationships between objects can be represented as a graph, i.e., as a collection of nodes and edges, where each edge connects two nodes. In such a graph, the nodes represent objects in the system and the edges represent relationships between the objects.

[0003] For example, in computer science, graphs can be used to represent communication networks. Specifically, a website can be represented as a graph of nodes and edges, where each node represents a web page, and each edge connecting two nodes represents a link between the two web pages represented by the nodes in that pair. As another example, in chemistry, the chemical structure of a molecule can be represented as a graph, where each node represents an atom, and each edge between two nodes represents a chemical bond between the two atoms.

[0004] Once a system is represented as a graph, it is possible to obtain more information about the system by manipulating the graph in various ways, such as graph transformations, graph coloring, graph decomposition, etc. Also, once represented as a graph, systems can be easily compared by their similarity to the graphs of other systems.

[0005] Interestingly, the more similar two graphs are (i.e., the more it is possible to transform one graph into the other with little or no change in the relative positions of the nodes and edges), the more properties the systems they represent will share. In particular, it may be possible to infer from the solution of a problem in one graph the solution to the same problem in the other graph, and vice versa. For example, if two molecules can be represented by two similar graphs, knowledge of the reactivity of one molecule may allow us to predict the reactivity of the other molecule.

[0006] Mathematically, if two graphs are isomorphic, they are "perfectly similar". However, if you try to calculate whether two graphs are isomorphic using direct methods (the so-called "brute force" method), the time required grows exponentially with the number of nodes in the graph, making it an impractical amount of time.

[0007] It is also known to define the similarity between graphs as a number between 0 and 1, which quantifies how similar two graphs are to each other. That is, if a pair of graphs has graph isomorphism (i.e., the graphs are "completely similar"), their similarity will be 1, and even if the pair of graphs is not completely similar, their similarity can be quantified. This is useful, for example, in classifying the graphs into categories to which they belong based on a reference graph.

[0008] In light of these circumstances, there is growing interest in finding a way to estimate the similarity between two graphs without using direct methods to determine whether they are isomorphic. This can be achieved by using a graph theory tool called a graph kernel. As disclosed in a paper by VISHWANATHAN et al. (Non-Patent Document 1), a graph kernel is a function that estimates the similarity between two graphs.

[0009] In fact, graph kernels are commonly used in machine learning techniques such as support vector machines (SVM), cluster analysis, principal component analysis (PCA), and regression analysis to handle large amounts of data represented as graphs and perform various tasks. As an example, to perform a classification task, a set of reference molecules can be classified into two types, reactive and non-reactive, with respect to their reactivity to a certain molecule. If the reactivity of the reference molecules is known, the performance of the graph kernel can be verified by analyzing how the reference molecules are classified using the classification kernel against the classification expected from experimental knowledge of the actual reactivity of known molecules.

[0010] More generally, the performance of a given graph kernel can be evaluated for its accuracy by using it to apply machine learning techniques to several pairs of reference graphs (called "benchmark data sets") that are representative of known systems, and comparing the results to those that would be obtained from perfect knowledge of the benchmark data set. Thus, one graph kernel is more accurate than another if the former is better suited to estimating the similarity between two graphs.

[0011] In particular, when a graph kernel is applied to a classification task, the accuracy of the graph kernel can be quantified by an F-score, for example, a number between 0 and 1 (or an equivalent percentage between 0% and 100%). Examples of datasets include AIDS, ENZYMES, FINGERPRINT, and PROTEINS, but in general, the F-score of a kernel depends on the benchmark dataset used. Therefore, there is an interest in finding the highest F-score, especially a graph kernel that can be applied to the most benchmark datasets and show the highest F-score.

[0012] In general, a graph kernel is a process that calculates two feature vectors of two graphs in the first step, and estimates the similarity between the two graphs based on the two feature vectors in the second step. A feature vector is an object (usually a histogram of numerical values) that describes the structural properties of a graph. That is, the feature vector represents the characteristics of a graph and can be used to compare the graph with other graphs. The above first and second steps may be implemented in different ways.

[0013] For example, the first step of the graph kernel can consist of computing two feature vectors by counting the number of edges connected to each node in the two graphs. However, such classical graph kernels do not provide optimal estimates of the similarity between graphs and are not accurate for all benchmark datasets.

[0014] In this regard, Verdon et al. (Non-Patent Document 2) and Szegedy (Non-Patent Document 3) disclose quantum algorithms, such as a quantum neural network algorithm and a quantum approximate optimization algorithm, suitable for graph isomorphism classification, respectively.

[0015] Such algorithms belong to the category of digital processing techniques that are implemented by breaking down the algorithm into a series of quantum logic gates to form quantum circuits and implementing them on a quantum processor, but such digital processing techniques are known to take a long time to implement as the number of nodes in the graph increases.

[0016] On the other hand, Schuld et al. (Non-Patent Document 4) have proposed an analog quantum processing method that analyzes the similarity between two graphs. In the analog quantum processing method, the parameters of the quantum hardware are adjusted so that a solution to a given problem is brought about by reproducing a given Hamiltonian by the dynamics of the quantum hardware (for example, see Non-Patent Document 5, which describes the difference between analog quantum processing methods and digital quantum processing methods).

[0017] In the method disclosed in Schuld et al.'s paper (Non-Patent Document 4), quantum hardware known as a Gaussian Boson sampler, which consists of a light source emitting a beam of light, an interferometer, and a photon detector, is used to construct a feature vector that maps the graph, thereby providing a graph kernel that shows better accuracy than classical graph kernels on certain benchmark datasets. [Prior art documents] [Non-patent literature]

[0018] [Non-Patent Document 1] VISHWANATHAN, S. Vichy N., SCHRAUDOLPH, Nicol N., KONDOR, Risi, et al. “Graph kernels”. Journal of Machine Learning Research, 2010, vol. 11, p. 1201-1242 [Non-Patent Document 2] VERDON, Guillaume, MCCOURT, Trevor, LUZHNICA, Enxhell, et al. Quantum graph neural networks. arXiv preprint arXiv:1909.12264, 2019 [Non-Patent Document 3] SZEGEDY, Mario. What do QAOA energies reveal about graphs?. arXiv preprint arXiv:1912.12277, 2019 [Non-Patent Document 4] SCHULD, Maria, BRADLER, Kamil, ISRAEL, Robert, et al. “Measuring the similarity of graphs with a Gaussian boson sampler”. Physical Review A, 2020, vol. 101, no 3, p. 032314 [Non-Patent Document 5] HENRIET, Loic, BEGUIN, Lucas, SIGNOLES, Adrien, et al. Quantum computing with neutral atoms. Quantum, 2020, vol. 4, p. 327 Summary of the Invention [Problem to be solved by the invention]

[0019] However, the proposed quantum hardware is designed specifically for graphs with a given number of nodes and cannot be easily adapted to graphs with a different number of nodes. Moreover, the size of the interferometer grows exponentially as the number of nodes in the graph increases, so this configuration cannot accommodate graphs with a large number of nodes, e.g., more than 10. Furthermore, the accuracy of the graph kernels described above still needs improvement.

[0020] Therefore, there is a need for an alternative, analogous, quantum processing approach to measure similarity between graphs that can handle graphs with various numbers of nodes, including graphs with large numbers of nodes. [Means for solving the problem]

[0021] Hereinafter, the terms "consisting of" and "comprising" are non-limiting open-ended terms synonymous with (the same meaning as) "including" and "having," and do not exclude other components not specified. In addition, the range of the terms "about" and "substantially" when referring to a numerical value in this disclosure is synonymous with (the same meaning as) a range of 80% to 120%, preferably 90% to 110%, of the numerical value.

[0022] A first aspect of the present disclosure relates to a method for measuring similarity between two graphs, the method comprising: determining a feature vector for each of the two graphs; and calculating an estimate of the similarity between the two graphs by comparing the two feature vectors. Determining the feature vector for each of the two graphs involves a repeatable measurement process including the following steps: Using optical trapping techniques, providing an array of atoms (atomic array) arranged to model said graph, where each atom of said array corresponds to a node of said graph, where a possible interaction between two of said atoms corresponds to an edge of said graph, and where each atom of said array has a number of energy levels corresponding to a number of atomic states, including at least a first base state and at least a first excited state. Preparing the atoms of the array to be in at least the first reference state. applying an interaction sequence to the array, the interaction sequence including the sub-steps of exciting at least one atom from the at least a first reference state to the at least a first excited state by transmitting, with an electromagnetic source, at least a first electromagnetic pulse having a first energy to the array for a first excitation time, and further allowing the atom to freely evolve for a first evolution time. Calculating values ​​of observables of the atomic array by detecting at least atoms in the array in at least the first excited state after the interaction sequence with a photodetector. and repeating the repeatable measurement process a plurality of times to obtain a plurality of values ​​of the observable. A processing unit calculates the feature vector from the distribution of the plurality of numerical values ​​of the observables.

[0023] According to the method of the first aspect, a graph can be encoded by the evolution (as a function of time) of the quantum state of an atomic array. That is, by measuring the quantum state of the atoms of the array after a certain evolution time, a numerical value of a physical quantity (an observable) of the atomic array is obtained, and this observable is a feature of the graph. Thus, by measuring multiple values ​​of this observable, a feature vector can be constructed that represents the features of the graph. Thus, by calculating this feature vector for two different graphs, it is possible to measure the similarity between the two graphs.

[0024] The method according to the first aspect relies on a quantum system of atomic arrays that can be adapted to different shapes, and can contain different numbers of atoms. Therefore, the method can provide feature vectors for graphs with different shapes and different numbers of nodes, especially for graphs with more than 10 or even 50 nodes. The computation time of the method increases polynomially with the number of nodes in the graph, so that the method can be applied to graphs with up to 10,000 nodes. The main constraint of the method is the maximum number of atoms in the atomic array used for the implementation.

[0025] In general, a graph is characterized by its structural properties, i.e., the arrangement of the nodes and edges of the graph. In a given method, how much of the structural properties of the graph can be incorporated into the feature vector is also called the "representativeness" of the method for computing the feature vector from the graph. In general, a method for computing the feature vector from a graph with high expressiveness can open the way to estimate the similarity between graphs with a large number of structural properties.

[0026] In the method according to the first aspect, by using a quantum system to calculate the feature vector, many structural properties of the graph can be incorporated into the feature vector, thereby providing better representability than existing methods.

[0027] The method according to the first aspect may be referred to as a graph kernel, in particular a quantum graph kernel.

[0028] The increased expressiveness of the feature vector calculation due to the reliance on atomic arrays representing graphs allows the method according to the first aspect, when applied to classification tasks, to measure the similarity between two graphs with higher accuracy than existing methods, especially in terms of F-score. In particular, the measured similarity can be used to make better predictions about a graph based on its similarity to one or more reference graphs. For example, the predictions may be the reactivity of a molecule compared to the reactivity of known molecules.

[0029] Furthermore, the method according to the first aspect can be applied to the comparison between graphs containing a large number of nodes, in particular more than 10 nodes, advantageously more than 50 nodes.

[0030] In one or more embodiments, the comparison method between the two feature vectors includes using a comparison function that calculates the Jensen-Shannon distance.

[0031] In the detection, atoms in the at least first reference state can be detected, or atoms in the at least first excited state can be detected, or, to increase the reliability of the measurement, both atoms in the at least first reference state and atoms in the at least first excited state can be detected.

[0032] From detection of the atoms in at least a first excited state and at least a second reference state, various observables can be calculated.

[0033] In one or more embodiments, the observable quantity of the atomic array is at least one selected from the number of atoms in the first excited state, the energy of the atomic array, and the degree of correlation between the atoms constituting the atomic array.

[0034] In one or more embodiments, the repeatable measurement process is repeated multiple times with the same first evolution time.

[0035] By repeating the repeatable measurement process while keeping the first evolution time the same, multiple numerical values ​​can be obtained for the observables to form a probability distribution, which can then be used to construct a feature vector.

[0036] It is also possible to obtain an "average", i.e. the statistical mean value of the observable for the same evolution time, by averaging these values ​​of the observable obtained over the same evolution time.

[0037] In one or more embodiments, the repeatable measurement process is repeated multiple times with different first evolution times.

[0038] The repeatable measurement process can be repeated for different first evolution times to obtain multiple average values ​​for the observables, each corresponding to a different first evolution time. Using multiple average values ​​for such observables, a suitable distribution can be obtained for forming a feature vector.

[0039] In one or more embodiments, the feature vector is determined based on a distribution of a plurality of means of the observables.

[0040] In one or more embodiments, the first energy and the first excitation time are selected using machine learning techniques.

[0041] For example, the first energy and the first excitation time are optimized to improve the expressiveness of the feature vector, and the first energy and the first excitation time can be optimized to improve an F-score obtained when the method according to the first aspect is applied to a graph kernel used for a classification task.

[0042] In one or more embodiments, the first energy is varied during the first excitation time.

[0043] In one or more embodiments, the interaction sequence further comprises: after allowing the atoms to freely evolve for the first evolution time, exciting at least one atom of the array from the at least first base state to the at least first excited state by transmitting a second electromagnetic pulse having a second energy to the array for a second excitation time; allowing the atoms to freely evolve for a second evolution time; Includes.

[0044] Such embodiments provide a more expressive way of computing feature vectors for graphs, which in particular can be used to form graph kernels suitable for implementing machine learning techniques on more complex tasks.

[0045] In one or more embodiments, the repeatable measurement process is repeated multiple times with the same first evolution time and with the same second evolution time.

[0046] In one or more embodiments, the repeatable measurement process is repeated multiple times with varying first evolution times and / or varying second evolution times.

[0047] In one or more embodiments, calculating the feature vector comprises calculating a probability distribution from a Fourier transform of a distribution of a plurality of the observables for a plurality of the first evolution times.

[0048] In one or more embodiments, the at least a first base state is a hyperfine ground state, the at least a first excited state is a Rydberg state |a1>=|nS>, where n is about 40 to about 90, and the electromagnetic pulse is a laser beam.

[0049] In one or more embodiments, the at least first reference state and the at least first excited state are two dipole-coupled Rydberg states |0>=|nS> and |1>=|nP>, respectively, where n is about 40 to about 90, and the electromagnetic pulse is microwave light.

[0050] In one or more embodiments, the at least a first reference state and the at least a first excited state are selected such that the evolution of the observable of the atomic array over time is related to an Ising or XY Hamiltonian, more precisely, this means that such evolution of the observable is a function that depends on the Ising or XY Hamiltonian.

[0051] In one or more embodiments, the first energy, the first excitation time, the second energy and the second excitation time are selected using machine learning techniques.

[0052] For example, the first energy, the first excitation time, the second energy, and the second excitation time are optimized to improve the expressiveness of the feature vector, and the first energy, the first excitation time, the second energy, and the second excitation time can be optimized to improve an F-score obtained when applying the method according to the first aspect to a graph kernel used for a classification task.

[0053] In one or more embodiments, the first energy is varied during the first excitation time and / or the second energy is varied during the second excitation time.

[0054] In one or more embodiments, at least one edge between two nodes in the graph has a weight. In the method of the first aspect, the weight of an edge connecting two nodes in the graph is modeled as the strength of the interaction between the two atoms modeling those two nodes.

[0055] In one or more embodiments, at least one node in the graph has a weight. In the method according to the first aspect, the weight of a node in the graph is modeled by applying a specific detuning to the atoms that model the node with the weight. Thus, a graph of nodes with different weights can be modeled by atoms with different detuning.

[0056] In one or more embodiments, the comparison between the two feature vectors includes using a comparison function that calculates the Jensen-Shannon distance.

[0057] In a second aspect, the present disclosure relates to a method for classifying input data with reference to known data belonging to at least one category, the method comprising: preparing an input graph for the input data; providing at least one reference graph for the known data; Measuring similarity between graphs 1. calculating an estimate of similarity between the input graph and the at least one reference graph according to a method according to an embodiment; associating the input data with the at least one category based on the estimated similarity; Includes.

[0058] According to the method of the second aspect, multiple estimates of the similarity between the input graph and multiple reference graphs can be calculated. The multiple similarity estimates can then be averaged to obtain an average similarity. Averaging the similarities with different reference graphs can improve the accuracy of the method for classifying the input graph.

[0059] In a third aspect, the present disclosure relates to a system configured to carry out the method according to the first aspect, the system comprising: a population of atoms arranged at predetermined locations using optical trapping techniques, said arrangement modeling a graph, each atom of the array having a plurality of energy levels corresponding to a plurality of states including at least a first reference state and at least a first excited state, said atoms being prepared in said at least first reference state; an electromagnetic wave source configured to generate the at least a first electromagnetic wave pulse to excite at least one atom of the array from the at least a first reference state to the at least a first excited state; a photodetector configured to detect atoms in the at least first reference state or the at least first excited state after the at least first electromagnetic pulse; a processing unit configured to calculate the feature vector based on the atoms detected in the at least first reference state or the at least first excited state after the atoms have freely evolved for a first evolution time following the first electromagnetic pulse; Equipped with.

[0060] In the system according to the third aspect, the atomic ensemble can be reconfigured to accommodate different atomic arrays with different numbers of atoms and different atomic configurations. Therefore, the system according to the fourth aspect is suitable for calculating feature vectors of different graphs with different numbers of nodes without changing the equipment used to capture the atoms. Therefore, the system according to the third aspect is an improvement over the systems of the prior art, especially over systems consisting of quantum systems with interferometers designed for graphs with a fixed number of nodes.

[0061] In one or more embodiments, the electromagnetic source is a laser or a radio frequency generator. In one or more embodiments, the atom is a neutral atom.

[0062] Other advantages and features of the present invention will become apparent from the following detailed description of the invention, taken in conjunction with the accompanying drawings. [Brief description of the drawings]

[0063] [Figure 1] FIG. 2 illustrates a method for computing a feature vector of a graph in the present disclosure. [Diagram 2] FIG. 2 illustrates steps in a method for calculating a feature vector of a graph. [Diagram 3] FIG. 1 illustrates a method for computing a feature vector of a graph, which comprises a layered interaction sequence. [Figure 4] FIG. 2 illustrates a method for calculating similarity between two graphs in the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0064] FIG. 1 is a diagram illustrating an example of a method for calculating a feature vector f(G1) of a graph G1 herein. In a first step 11 of the method, an atomic array is prepared to model a graph G1. In this step 11, the array is populated with a number of atoms equal to the number of nodes of the graph G1. The atomic array may be, for example, neutral atoms in a quantum register stabilized in place using optical trapping techniques. The positions of the atoms and the distances between them are selected such that the structure formed by them models the nodes and edges of the graph. In particular, the atoms are arranged in the array such that two atoms representing two connected nodes can interact in the array, for example by dipole interactions.

[0065] In one or more embodiments, the graph being modeled has edges with weights, and the interactions between the atoms corresponding to those edges are also modeled with the weights, e.g., based on differences in the strength of the interactions between the atoms.

[0066] In one or more embodiments, the graph being modeled has weighted nodes, and the relative weights between nodes can be recreated, for example, by applying relative detuning to the atoms that model these weighted nodes.

[0067] Indeed, by using known trapping techniques and relocation algorithms, it is possible to arrange the atoms according to a predefined position (see, for example, BARREDO, Daniel, DE LESELEUC, Sylvain, LIENHARD, Vincent, et al. An atom-by-atom assembler of defect-free arbitrary two-dimensional atomic arrays. Science, 2016, vol. 354, no 6315, p. 1021-1023). In particular, the position of the atoms can be changed by changing the position of the optical trap in the array, for example by means of a spatial light modulator. In such an array, the distance between two connected atoms may be about 4 to 150 micrometers.

[0068] The atoms constituting the atomic array have multiple energy levels corresponding to multiple atomic states. Each atom in the array has at least two states including a base state (a0) and an excited state (a1). However, the atoms may also exhibit multiple base states and / or multiple excited states.

[0069] In one or more embodiments, the atoms are Rydberg atoms having a hyperfine structure exhibiting a number of degenerate hyperfine levels, the degeneracy being resolvable by the application of a static magnetic field, and the base state and the excited state can be selected from among the hyperfine levels.

[0070] In a second step 13, the atoms of the atomic array are prepared in a reference state a0. This means that each atom of the array is prepared in a reference state a0. The reference state a0 of the atoms of the array may be a ground state. However, it may also be a state other than the ground state, such as a Rydberg state. When using a reference state corresponding to a degenerate hyperfine level, it may be possible to remove the degeneracy by applying an external static magnetism to the atomic array.

[0071] After step 13, an interaction sequence is applied to the atomic array, including steps 14 and 15.

[0072] In step 14, a first electromagnetic pulse EP1 is sent to the atomic array for a first excitation time Tex1, thereby exciting at least one of the atoms from the base state to a predefined excited state.

[0073] In order to excite the atoms between the reference state and the excited state, the frequency of the electromagnetic pulse EP1 is set to a frequency corresponding to energy approximately equal to the energy difference between the energy of the reference state and the energy of the excited state.

[0074] Such excitation of atoms between two states, consisting of a reference state and an excited state, is also called Rabi oscillation. During the Rabi oscillation, the probability of finding an atom in an excited state in the array varies as a function of the first excitation time Tex1. The amplitude and frequency of the variation depend on several parameters, including the energy of the electromagnetic pulse and the amplitude of the electromagnetic pulse. In practice, the parameters are selected such that after the first excitation time, at least one atom of the array is in an excited state. Also, the parameters can be selected such that after the first excitation time, at least two or more atoms of the array are in an excited state.

[0075] Typically, the first excitation time Tex1 is between 4 nanoseconds and 10 microseconds.

[0076] In step 15, pulse EP1 is interrupted and the atoms are allowed to evolve freely for a first evolution time (T1), where "freely" means that no electromagnetic pulses are sent to the atomic array.

[0077] In step 17, the states of the atoms in the array are measured. Based on this measurement, an observable quantity O1 of the quantum system formed by the atomic array is determined.

[0078] Examples of observables include the total number of atoms in an excited state, the energy of an atomic array, the correlation function of atom pairs (a numerical representation of whether or not an atom pair is detected in an excited state simultaneously), and, more broadly, any physical quantity that can be measured for a quantum state consisting of an atomic array.

[0079] Indeed, when measuring the state of the atoms of the array, each atom is detected either in a reference state or in an excited state, which allows one to obtain a list of the states of the atoms in the array, for example a list of 0s and 1s, i.e. a bit string, where each element of the bit string corresponds to a state of an atom in the array (for example a 0 if the atom is in the reference state and a 1 if the atom is in the excited state).

[0080] After step 17, steps 11, 13, 14, 15, and 17 are repeated multiple times (N times) to obtain multiple values ​​for the observables {O1_i}_(1≦i≦N). Therefore, in this specification, steps 11, 13, 14, 15, and 17 are collectively referred to as the repeatable measurement process. When the repeatable measurement process is completed, the atoms in the atomic array are discarded, and a new atomic array is provided when starting an iteration of the repeatable measurement process (step 11).

[0081] Due to the probabilistic nature of measurements on a quantum system (in this example the array of atoms), the measured state for each atom may be different at each iteration of the repeatable measurement process.

[0082] Thus, the repeatable measurement process may be repeated multiple times with the same first evolution time T1, resulting in multiple values ​​of the observable O1 being obtained to construct a probability distribution, which the applicant has determined can be used to directly form a feature vector for the graph G1 modeled with the atomic array.

[0083] In another embodiment, the repeatable measurement process is repeated multiple times by varying the first evolution time {T1_j} (1≦j≦M).

[0084] In this case, the applicant of the present application obtains a plurality of values ​​of the observables {O1(T1)_i}_(1≦i≦N) for each first evolution time T1, and performs a statistical average to obtain an average value O1 of the observables. ave It was confirmed that (T1) was obtained.

[0085] Therefore, the mean value of the observables for each different first evolution time (T1_j) is O1 ave By calculating ave (T1_j)}_(1≦j≦M).

[0086] Applicants have shown that a probability distribution can be calculated using multiple averages of observables to form a feature vector.

[0087] Specifically, applicants have determined that the Fourier transform of multiple averages of observables over multiple first evolution times forms a probability distribution that can be used as a feature vector for a graph (G1) modeled with an atomic array.

[0088] In one or more embodiments, the total number of iterations is selected in relation to the desired number N for obtaining an average value for each first evolution time T1 and the desired number M of different first evolution times to be considered. In particular, since several iterations are performed to obtain a statistical average of observables at the same first evolution time, the number M of different first evolution times will generally be lower than the total number N of iterations.

[0089] FIG. 2 is a schematic diagram illustrating the process of computing a feature vector f(G1) of an input graph (graph G1) in the method described herein.

[0090] In the example of FIG. 2, graph G1 is a graph representing graph-structured data (e.g., a social network, a database, etc.) and consists of four nodes (such as nodes 111 and 115) and four edges (such as edge 113), each connecting two of the four nodes.

[0091] Next, an array 120 of atoms is prepared to model graph G1, where the arrangement of atoms in array 120 corresponds to graph G1. Specifically, in the example of Figure 2, atoms 121 model nodes 111 of graph G1, and interactions between atoms 121, 125, represented by links 123, model edges 113 of graph G1.

[0092] The atoms of the array 120 are prepared in a reference state a0. The reference state a0 of the atoms of the array may be a ground state, but may also be a state other than the ground state, such as a Rydberg state.

[0093] 2 also shows a plot 130 of the evolution over time of an exemplary observable (O1) of the atomic array 120. Such evolution over time goes through two phases, an excitation phase 131 and a free evolution phase 133, which constitute an interaction sequence 137. The free evolution phase 133 is also referred to herein as the evolution phase.

[0094] In other examples, an interaction sequence 137 may include multiple excitation phases 131 and / or multiple evolution phases 133 .

[0095] In the excitation phase 131, an electromagnetic pulse EP1 of energy E1 is applied to the array of atoms 120 for an excitation time Tex1. Such an electromagnetic pulse EP1 makes it possible to cause the atoms to oscillate between a reference state a0 and a given excited state a1, also commonly called Rabi oscillations. During this oscillation, the probability of finding each atom of the array in an excited state oscillates between a maximum and a minimum at a given frequency Ω, called the Rabi frequency. The Rabi frequency depends, inter alia, on the combination of the two states in which the atom oscillates and on the energy E1 of the electromagnetic pulse EP1.

[0096] In particular, the excitation time Tex1 and energy E1 can be selected to result in a predetermined number of atoms in the array evolving from a base state a0 to a predefined excited state a1. Typically, the excitation of at least one atom from a base state to an excited state is sufficient for the method described herein to calculate a feature vector.

[0097] In the excitation phase 131, the state evolution of the atoms is mainly driven by the interaction Hamiltonian describing the interaction between the atoms and the electromagnetic pulse, while the effect of the intrinsic Hamiltonian of the atomic array, which is the Hamiltonian describing the evolution of the atomic array when no external electromagnetic pulse is applied to the array, can be neglected.

[0098] In general, the energy of the electromagnetic pulse EP1 is approximately equal to the difference between the energy of the reference state a0 and the energy of the excited state a1. Due to this energy constraint, the average frequency of the electromagnetic pulse is also constrained.

[0099] When the electromagnetic pulse EP1 is discontinued, the excitation phase 131 ends and an evolution phase 133 begins, during which the atoms of the array 120 are free to evolve during an interaction time T1. During the evolution phase 133, the atoms of the array continue to evolve between a reference state a0 and an excited state a1, such evolution being determined by the intrinsic Hamiltonian of the array of atoms.

[0100] In fact, this suggests that during the evolution phase 133, the observables O1 of the array of atoms follow a function h(t) that has a specific relationship to the intrinsic Hamiltonian of the array of atoms, and therefore to the graph G1 modeled by the array of atoms 120.

[0101] After the evolution phase 133, the values ​​of observables O1 specific to graph G1 can be determined by measuring 135 the states of the atoms of the array.

[0102] Various observables can be considered, such as the total number of atoms in an excited state, the energy of the array of atoms, and the correlation function of atom pairs (which represents, in numerical terms, whether or not an atom pair is detected in the excited state at the same time).

[0103] The methods described herein can be performed with an array of atoms having a wide variety of base and excited states. Depending on the choice of atomic states included in the interaction sequence 137, the atoms in the array will experience different interactions and, as a result, different Hamiltonians.

[0104] For example, the hyperfine ground state may be selected as the reference state and the Rydberg state (eg, |1>=|nS>, where n is between 40 and 90) may be selected as the excited state.

[0105] With this selection of reference and excited states, each atom in the array behaves as a two-level system whose dynamics are determined by a specific Hamiltonian called the Ising Hamiltonian. In such a configuration, the energy difference between the reference and excited states typically corresponds to the energy of a photon in the visible light range. Thus, the electromagnetic pulse used in the excitation phase 131 is a laser beam or multiple laser beams (e.g., two counter-propagating beams with wavelengths of about 420 nm and about 1013 nm).

[0106] The Ising Hamiltonian is a function that is relevant to many problems in condensed matter. For example, in materials science, it describes how quantum magnets evolve at ultra-low temperatures. Interestingly, this function can also be used in a wide variety of classical applications, as it can give optimal solutions to computationally difficult problems.

[0107] As another example, we select two dipole-coupled Rydberg states (e.g., |0>=|nS> and |1>=|nP>, where n is between 40 and 90) for the base state |0> and the excited state |1>. With such a selection of reference and excited states, each atom of the array behaves as a two-level system whose dynamics are determined by a specific Hamiltonian, called the XY Hamiltonian. In such a configuration, the energy difference between the reference and excited states generally corresponds to the energy of a photon in the microwave range. That is, the electromagnetic pulse used in the excitation phase 131 will be microwave radiation generated by a radio frequency source.

[0108] If the selected reference or excited states are degenerate, a static magnetic field can be used to break the degeneracy.

[0109] In general, to choose a different Hamiltonian, one simply chooses a different set of states, and the energy (and therefore frequency) of the electromagnetic pulse used in the excitation phase is selected according to the properties of this set of states.

[0110] The choice of reference and excited states can also affect the method for determining the number of atoms in an excited state. For example, in the case of a detection method that is sensitive only to atoms in the Rydberg state, it is possible to directly measure the atoms in an excited state by using a combination of states that corresponds to the Ising Hamiltonian. However, in the case of a combination of Rydberg states that corresponds to the XY Hamiltonian, an additional step may be applied in which atoms in one of the two states can be removed from the array and the remaining atoms in the array can be counted.

[0111] Figure 3 shows an example of a method for calculating a feature vector, which includes a layered interaction sequence 137. Compared with the method shown in Figure 2, the layered interaction sequence 137 has two layers 144, 146. Each layer 144, 146 consists of an excitation phase 131, 141 and an evolution phase 133, 143, respectively.

[0112] The excitation phase 141 of the second layer 146 includes a step of oscillating the atoms between a reference state a0 and a predefined excited state a1 by sending a second electromagnetic pulse EP2 (of energy E2) to the atomic array for a second excitation time Tex2. Specifically, the second excitation time Tex2 and the second energy E2 are selected such that at least one or more atoms of the array evolve from the reference state a0 to a predefined excited state a1. In this excitation phase 141, the state evolution of the atoms is mainly driven by an interaction Hamiltonian describing the interaction between the array of atoms and the second electromagnetic pulse EP2, while the effect of the intrinsic Hamiltonian of the array of atoms can be neglected.

[0113] In the second layer 146, similar to the first layer 144, after the excitation phase 141 ends an evolution phase 143 begins, where the atoms of the array are free to evolve for a second evolution time T2.

[0114] The process after the second layer 146 is similar to method 10 described above in FIG. 1, except that the second evolution time T2 can be changed along with or apart from the first evolution time T1 between each iteration.

[0115] This allows obtaining a distribution of the observables {O1_i} for the first and second evolution times {T1_i, T2_i} after multiple iterations of the repeatable measurement process. The number of iterations can be selected in relation to the desired number of averages for each of the first and second evolution times {T1_i, T2_i} of interest and the desired number of changes of the first and second evolution times.

[0116] In other examples not shown, the layered interaction sequence 137 can have multiple layers, for example including 3 layers or up to 10 layers. Increasing the number of layers can increase the expressiveness of the method and provide feature vectors suitable for better similarity estimation. In particular, when the method is used with a graph kernel for classification tasks, increasing the number of layers can increase the F-score of the graph kernel for classification.

[0117] FIG. 4 illustrates a method 30 for measuring the similarity between two graphs G1, G2 as used herein.

[0118] In the method 30, a first feature vector f(G1) for graph G1 is calculated using the method 10 shown in FIG. 1 or any of its variations, and a second feature vector f(G2) for graph G2 is similarly calculated. These two feature vectors f(G1), f(G2) are then compared (20) using a comparison function to obtain an estimate of the similarity K(G1,G2) between the two graphs G1, G2. The method 30 can be referred to as a graph kernel since it estimates the similarity between the two graphs G1, G2. The method 30 can be referred to as a quantum graph kernel since it calculates the feature vectors through a quantum system (atomic array 120).

[0119] Method 30 may use a variety of functions for comparison 20, such as, for example, a linear kernel, a radial basis function (RBF) kernel, a function that calculates the Jensen-Shannon distance, etc. Although such comparison functions may have names that include the term "kernel," they should not be confused with the "graph kernel" of method 30 herein. Specifically, the input of such comparison functions is a feature vector, whereas the input of a graph kernel is a graph.

[0120] While the present invention has been described in connection with certain embodiments, those skilled in the art having access to this disclosure will recognize that other embodiments may be made without departing from the spirit of the invention disclosed herein, and therefore, the scope of the present invention should be limited only by the scope of the appended claims. The present invention includes the following embodiments. [Aspect 1] A method (30) for measuring the similarity (K) between two graphs (G1, G2), , a step (10) of determining feature vectors (f1, f2) for each of the two graphs (G1, G2); and calculating (20) an estimate of the similarity (K) between the two graphs (G1, G2) by comparing the two feature vectors (f1, f2), The process of obtaining the feature vector for each of the two graphs (G1, G2) includes the following steps: an array of atoms arranged to model said graph (G1), with each atom (121) corresponding to a node (111) of said graph (G1) and with each interaction (123) between two atoms corresponding to an edge (113) of said graph (G1), an atomic array (120) in which each atom (121) has a plurality of energy levels corresponding to a plurality of atomic states including at least a first reference state (a0) and at least a first excited state (a1); Providing by optical trapping techniques; - preparing the atoms of the array to be in the at least first reference state (a0); exciting at least one atom of the array from said at least a first base state (a0) to said at least a first excited state (a1) by transmitting at least a first electromagnetic pulse (EP1) having a first energy (E1) to said array using an electromagnetic source for a first excitation time (Tex1); and allowing the atoms to freely evolve for a first evolution time (T1). applying an interaction sequence (137) to the array; , determining a value of an observable (O1) of the atomic array by detecting, by means of a photodetector, at least atoms in the array that are in the at least first excited state (a1) after the interaction sequence (137);

[0033] A repeatable measurement process including Repeating the repeatable measurement process a number of times to obtain a number of values ​​of the observable (O1); A processing unit calculates the feature vector (f) from the distribution of the plurality of numerical values ​​of the observable quantity (O1); method. [Aspect 2] A method according to aspect 1, wherein the repeatable measurement process is repeated multiple times with the first evolution time (T1) being the same. [Aspect 3] 3. The method according to claim 1 or 2, wherein the repeatable measurement process is repeated multiple times by varying the first evolution time (T1). Aspect 4 The method of any one of the preceding aspects, wherein the first energy (E1) is varied during the first excitation time (Tex1). Aspect 5 In the method according to aspect 1, the interaction sequence (137) further comprises: after allowing the atoms to freely evolve for the first evolution time (T1), exciting at least one atom of the array from the at least a first base state (a0) to the at least a first excited state (a1) by sending a second electromagnetic pulse (EP2) having a second energy (E2) to the array for a second excitation time (Tex2); allowing said atoms to freely evolve for a second evolution time (T2); A method comprising: Aspect 6 A method according to embodiment 5, wherein the repeatable measurement process is repeated multiple times with the same first evolution time (T1) and the same second evolution time (T2). Aspect 7 A method according to embodiment 5 or 6, wherein the repeatable measurement process is repeated multiple times by varying the first evolution time (T1) and / or by varying the second evolution time (T2). Aspect 8 The method according to any one of aspects 5 to 7, wherein the first energy (E1) is varied during the first excitation time (Tex1) and / or the second energy (E2) is varied during the second excitation time (Tex2). Aspect 9 9. The method of any one of aspects 1 to 8, wherein the calculation of the feature vector comprises calculation of a probability distribution from a Fourier transform of a distribution of the plurality of numerical values ​​of the observable (O1) for a plurality of the first evolution times. Aspect 10 In the method according to any one of aspects 1 to 9, the observable quantity (O1) of the atomic array is one selected from the number of atoms in the at least first excited state, the energy of the atomic array, and the degree of correlation between atoms in the atomic array. Aspect 11 11. The method according to any one of aspects 1 to 10, wherein the at least a first base state (a0) is a hyperfine ground state, the at least a first excited state (a1) is a Rydberg state |a1>=|nS>, where n is about 40 to about 90, and the electromagnetic wave pulse is a laser beam. Aspect 12 12. The method according to any one of aspects 1 to 11, wherein the at least a first base state (a0) and the at least a first excited state (a1) are two dipole-coupled Rydberg states of |0>=|nS> and |1>=|nP>, respectively, where n is about 40 to about 90, and the electromagnetic wave pulse is microwave light. Aspect 13 A system configured to carry out the method (30) according to any one of aspects 1 to 12, comprising: a population of atoms arranged at predetermined locations using optical trapping techniques, said arrangement modeling a graph, each atom of the array exhibiting a plurality of energy levels corresponding to a plurality of states including at least a first reference state (a0) and at least a first excited state (a1), said atoms being prepared in said at least first reference state; an electromagnetic wave source configured to generate the at least a first electromagnetic wave pulse (EP1) to excite at least one atom of the array from the at least a first reference state (a0) to the at least a first excited state (a1); a photodetector configured to detect atoms in the at least first reference state (a0) or the at least first excited state (a1) after the at least first electromagnetic pulse (EP1); a processing unit configured to calculate the feature vector based on the atoms detected in the at least first reference state or the at least first excited state after the atoms have freely evolved for a first evolution time following the first electromagnetic pulse; A system comprising: [Explanation of symbols]

[0121] G1, G2 graphs f(G1), f(G2) feature vectors 120 Atomic Array 121, 125 atoms 123 Link a0 Reference condition a1 excited state Tex1 First excitation time Tex2 Second excitation time EP1, EP2 Electromagnetic Pulse T1 First evolution time T2 Second evolution time O1 Observables 111, 115 nodes 113 Edge

Claims

1. A method (30) for measuring a similarity (K) between two graphs (G1, G2), comprising the steps of: determining (10) a feature vector for each of the two graphs (G1, G2); and calculating (20) an estimate of the similarity (K) between the two graphs (G1, G2) by comparing the two feature vectors; The process of obtaining the feature vector for each of the two graphs (G1, G2) includes the steps of: an array of atoms arranged to model said graph (G1), with each atom (121) corresponding to a node (111) of said graph (G1) and with each interaction (123) between two atoms corresponding to an edge (113) of said graph (G1), an atomic array (120) in which each atom (121) has a plurality of energy levels corresponding to a plurality of atomic states including at least a first reference state (a0) and at least a first excited state (a1); Providing by optical trapping techniques; - preparing the atoms of the atomic array to be in the at least first reference state (a0); - exciting at least one atom of the atomic array from the at least a first base state (a0) to the at least a first excited state (a1) by transmitting at least a first electromagnetic pulse (EP1) having a first energy (E1) to the atomic array using an electromagnetic source for a first excitation time (Tex1); and allowing the atoms to freely evolve for a first evolution time (T1). applying an interaction sequence (137) to the atomic array; - calculating a value of an observable (O1) of the atomic array by detecting, by means of a photodetector, atoms in the atomic array that are in at least the first excited state (a1) after the interaction sequence (137); [0033] A repeatable measurement process including Repeating the repeatable measurement process a number of times to obtain a number of values ​​of the observable (O1); A processing unit calculates the feature vector from the distribution of the plurality of numerical values ​​of the observable quantity (O1); method.

2. 2. The method of claim 1, wherein the repeatable measurement process is repeated multiple times with the first evolution time (T1) being the same.

3. 3. The method of claim 1 or 2, wherein the repeatable measurement process is repeated a number of times by varying the first evolution time (T1).

4. 3. The method according to claim 1 or 2, wherein the first energy (E1) is varied during the first excitation time (Tex1).

5. 2. The method of claim 1, wherein the interaction sequence (137) further comprises: after allowing the atoms to freely evolve during the first evolution time (T1), exciting at least one atom of the atomic array from the at least a first base state (a0) to the at least a first excited state (a1) by sending a second electromagnetic wave pulse (EP2) having a second energy (E2) to the atomic array for a second excitation time (Tex2); allowing the atoms to freely evolve for a second evolution time (T2); A method comprising:

6. 6. The method of claim 5, wherein the repeatable measurement process is repeated multiple times with the same first evolution time (T1) and with the same second evolution time (T2).

7. 7. The method according to claim 5 or 6, wherein the repeatable measurement process is repeated multiple times by varying the first evolution time (T1) and / or by varying the second evolution time (T2).

8. 7. The method according to claim 5 or 6, wherein the first energy (E1) is varied during the first excitation time (Tex1) and / or the second energy (E2) is varied during the second excitation time (Tex2).

9. 6. The method according to claim 1 or 5, wherein the calculation of the feature vector comprises the calculation of a probability distribution from a Fourier transform of the distribution of the plurality of numerical values ​​of the observable (O1) for a plurality of the first evolution times.

10. 6. The method according to claim 1 or 5, wherein the observable quantity (O1) of the atomic array is one selected from the number of atoms in the at least first excited state, the energy of the atomic array, and the degree of correlation between atoms in the atomic array.

11. 6. The method of claim 1 or 5, wherein the at least a first base state (a0) is a hyperfine ground state, the at least a first excited state (a1) is a Rydberg state |a1>=|nS>, where n is from about 40 to about 90, and the electromagnetic pulse is a laser beam.

12. 6. The method of claim 1 or 5, wherein the at least a first base state (a0) and the at least a first excited state (a1) are two dipole-coupled Rydberg states |0>=|nS>, |1>=|nP>, respectively, where n is about 40 to about 90, and the electromagnetic wave pulse is microwave light.

13. A system configured to carry out the method (30) according to claim 1 or 5, comprising: an atomic population arranged at a predetermined location using optical trapping techniques, said arrangement modeling a graph, each atom of said atomic array exhibiting a plurality of energy levels corresponding to a plurality of states including at least a first reference state (a0) and at least a first excited state (a1), said atoms being prepared in said at least first reference state; an electromagnetic wave source configured to generate the at least a first electromagnetic wave pulse (EP1) to excite at least one atom of the atomic array from the at least a first reference state (a0) to the at least a first excited state (a1); a photodetector configured to detect atoms in the at least first reference state (a0) or the at least first excited state (a1) after the at least first electromagnetic pulse (EP1); a processing unit configured to calculate the feature vector based on the atoms detected in the at least first reference state or the at least first excited state after the atoms have freely evolved for a first evolution time following the first electromagnetic pulse; A system comprising: