Quantum supremacy using quantum circuits for gradient estimation

A quantum circuit with a phase oracle using finite difference approximation addresses the limitation of existing algorithms by enabling efficient gradient estimation for a wider range of functions, offering a seven-fold reduction in computational resources for financial derivative calculations.

JP2024542226APending Publication Date: 2024-11-13GOLDMAN SACHS & CO LLC +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2024529773
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2021-11-19
Filing Date
2022-11-18
Publication Date
2024-11-13

AI Technical Summary

Technical Problem

Existing quantum gradient algorithms are limited to a specific class of smooth functions (Gevrey class G 1/2) and do not guarantee complexity scaling for functions that describe real-world phenomena, which are often not in this class.

Method used

A quantum circuit implementing a phase oracle using a finite difference approximation with an order greater than zero to determine gradients, allowing for improved gradient estimation on a broader range of functions without requiring them to be smooth.

Benefits of technology

The proposed quantum gradient algorithm provides advantages over classical computation by achieving complexity scaling of O(1/ε p ) for gradient estimation, enabling efficient calculation of gradients for financial derivatives and reducing computational resources by up to seven times compared to classical methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024542226000001_ABST
    Figure 2024542226000001_ABST
Patent Text Reader

Abstract

Described herein is a quantum gradient algorithm that results in quantum advantages over classical approaches. In an example, a quantum circuit is configured to implement the quantum gradient algorithm running on qubits of a quantum computing system. The quantum gradient algorithm is based on a phase oracle O defined by a finite difference approximation having an order greater than 0. sf m where the complexity of the quantum gradient algorithm scales as O(√k / ε). A quantum circuit is repeatedly run on the qubits of a quantum computing system to determine the k-dimensional gradient for a function f(x) at a point x0 within an error ε.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates generally to quantum gradient algorithms, and more particularly to quantum circuits that implement phase oracles to determine gradients. [Background technology]

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of U.S. Provisional Patent Application No. 63 / 281,313, filed November 19, 2021, which is incorporated herein by reference.

[0003] Non-Patent Document 1 and Non-Patent Document 2 describe quantum algorithms for determining the gradient of a function. However, the algorithms are limited to a certain class of smooth functions (the Gevrey class G 1 / 2 The advantages of complexity scaling over classical computation can only be guaranteed if applied to functions that are functions. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Jordan(arXiv:quant-ph / 0405146) [Non-Patent Document 2] Gilyen (arXiv:1711.00465) [Non-Patent Document 3] Chakrabarti (arXiv:2012.03819) [Non-Patent Document 4] arXiv:1610.06546v3 [Non-Patent Document 5] arXiv:1806.01838 [Non-Patent Document 6] Martyn (arXiv:2110.11327) Summary of the Invention

[0005] Many functions describing real-world phenomena belong to the Gevrey class G 1 / 2 Therefore, the algorithms described in [1] and [2] cannot guarantee the advantage of complexity scaling for these types of functions. In contrast, the present disclosure is based on the assumption that the function is of Gevrey class G 1 / 2 We describe a quantum circuit that implements a quantum gradient algorithm (called SFQG), which provides an advantage over classical computation of functions without requiring them to be functions.

[0006] In some embodiments, the quantum gradient algorithm (e.g., the SFQG method) uses a phase oracle defined by a finite difference approximation having an order greater than 0.

[0007]

number

[0008] The quantum gradient algorithm may be implemented by preparing and executing a quantum circuit on the qubits of a computing system.

[0009]

number

[0010] teeth,

[0011]

number

[0012] where m is the finite difference approximation order greater than 0, |x> is a k-dimensional vector representing a set of k parameters, and S is a scaling factor that controls the accuracy of the finite difference approximation.

[0013]

number

[0014] are the coefficients of a finite difference approximation of order m. For a function f(x) describing a quantity of a resource, where x represents a set of k parameters of this quantity, the quantum gradient algorithm can be repeatedly executed by a quantum computing system to determine the k-dimensional gradient of the function f(x) within error ε at a point x0. Among its advantages, the complexity of the quantum gradient algorithm is

[0015]

number

[0016] Scale as follows:

[0017] Other aspects include components, devices, systems, improvements, methods, processes, applications, computer readable media, and other technology related to any of the above.

[0018] Aspects of the present disclosure have other advantages and features that will become more readily apparent from the following detailed description and the appended claims, when considered in conjunction with the accompanying drawing examples. [Brief description of the drawings]

[0019] [Figure 1] FIG. 1 illustrates an exemplary probability distribution resulting from a numerical simulation of the quantum gradient algorithm. [Diagram 2] FIG. 13 illustrates an example probability distribution resulting from a second numerical simulation of the quantum gradient algorithm. [Diagram 3] FIG. 1 is an exemplary quantum circuit diagram for the Simulation-Free Quantum Gradient (SFQG) algorithm. [Figure 4]FIG. 13 illustrates an example probability distribution resulting from a numerical simulation of the SFQG algorithm. [Figure 5a] FIG. 5 illustrates an exemplary discrete probability distribution for the basket option of FIG. 4. [Figure 5b] FIG. 5 illustrates an exemplary log-likelihood plot for the basket option of FIG. 4. [Figure 6] FIG. 1 illustrates exemplary simulated measurement results. [Figure 7] 1 is a method chart illustrating an example method for determining the gradient of a function f(x) describing an amount of a resource. [Figure 8A] FIG. 1 is a block diagram illustrating a computing system in accordance with one or more aspects. [Figure 8B] FIG. 1 is a block diagram illustrating a computing system in accordance with one or more aspects. [Figure 8C] FIG. 1 is a block diagram illustrating an aspect of a quantum computing system. [Figure 8D] FIG. 1 is a block diagram illustrating an aspect of a quantum computing system. [Figure 8E] 1 is a flow chart illustrating an exemplary execution of a quantum routine on a computing system. [Figure 9] 1 is an exemplary architecture of a classical computing system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] The figures and the following description relate to preferred embodiments for purposes of illustration only. It should be noted from the following description that alternative aspects of the structures and methods disclosed herein are readily discernible as viable alternatives that may be employed without departing from the principles of what is claimed.

[0021] In particular, a quantum algorithm is disclosed for computing market risk for financial derivatives. Quantum amplitude estimation can accelerate derivative pricing quadratically in the target error, which can be extended to the advantage of squared error scaling in market risk computation. Employing a quantum gradient estimation algorithm can provide a further quadratic advantage in the number of relevant market sensitivities, called greeks. By numerically simulating the quantum gradient estimation algorithm for financial derivatives of practical interest, exemplary greeks can be successfully estimated, but the resource requirements can be substantially lower in practice than expected by the theoretical complexity limit. This further advantage in computation reduces the estimated logical clock rate used for quantum advantage from [3] by a factor of about 7, from 50 MHz to 7 MHz, even for a modest number of greeks according to the industry standard [4]. Furthermore, using a system with sufficient resources, the quantum algorithm can be parallelized across up to 60 QPUs, where the logical clock rate of each device used is about 100 kHz to achieve the same overall execution time as a serial run. Throughout this disclosure, several different combinations of quantum and classical approaches that may be used to calculate the market risk of financial derivatives (among other things) are summarized and compared.

[0022] 1. Introduction Quantum algorithms have the potential to accelerate several real-world computations, such as financial derivatives pricing and risk analysis. Some of these algorithms use quantum amplitude estimation to achieve second-order advantages over classical Monte Carlo methods, which are used in practice for most computationally expensive pricing. Let us define the error in pricing as ε p Quantum supremacy is

[0023]

number

[0024] The quantum algorithm comes from the fact that the running time of a classical Monte Carlo simulation scales with O(1 / ε p ).

[0025] A related important financial application is the calculation of the sensitivity of derivative prices to model and market parameters. This amounts to calculating the gradient of the derivative prices with respect to the input parameters. An exemplary business use of calculating these gradients is to enable hedging of market risk arising from exposure to derivative contracts. Hedging this risk can be crucial for financial institutions. In some cases, the gradient can be calculated analytically, for example when the derivative prices have an analytical form. In some aspects, the disclosed systems and methods calculate gradients when no closed form exists for the prices and classical comparisons become Monte Carlo simulations. Gradients of financial derivatives are commonly referred to as Greeks, since they are commonly labeled using letters of the Greek alphabet. For k-greeks (k-dimensional gradients), classical finite difference methods are

[0026]

number

[0027] For scaling of , we may calculate prices at multiple points in each parameter dimension.

[0028] One approach to quantum acceleration of the Greeks computation is to construct a finite-difference approximation of the price of a financial derivative on a quantum computer and perform amplitude estimation on that quantity instead of its price. Section 2.4 shows that this approach, called the semi-classical method, scales as (k / c), where c is the target error in the gradient. However, this approach can be further improved in scaling in the Greeks number k by using a quantum algorithm for computing the gradient.

[0029] Quantum algorithms for computing the components of the gradient in a superposition were originally introduced by [1] and more recently revisited by [2], but these references do not consider the application of quantum algorithms to Greek computations. [2] proposes that quantum algorithms can be applied to Greek computations in general with O(k / ε 2 ) to show that the algorithm scales as

[0030]

number

[0031] The quantum gradient algorithm of for a certain class of smooth functions

[0032]

number

[0033] For simplicity, in the remainder of this disclosure we refer to this result as the GAW quantum gradient algorithm.

[0034] In this disclosure, the algorithmic framework of [2] is applied to compute financial Greeks and to study several types of quantum gradient algorithms. The quantum pricing method of [3] is used as a subroutine of the GAW algorithm to numerically estimate the resource requirements of computing the Greeks of two types of option contracts: (a) European call options, which can be used as a benchmark to establish the validity of the algorithm and to estimate the corresponding resource requirements, and (b) path-dependent basket options, whose pricing profile is representative of typical financial derivative contracts of practical interest.

[0035] This disclosure also introduces a second-order exact quantum gradient algorithm, m=1 version, for which an explicit and compact quantum implementation in the case of financial derivatives is disclosed that does not rely on block coding or Hamiltonian simulation. This simulation-free quantum gradient (SFQG) approach can be used to compute the Greeks of path-dependent derivatives and can be constructed significantly cheaper than Hamiltonian-based methods. Furthermore, the disclosed approach can improve the overall performance of the quantum gradient estimation algorithm by employing a maximum likelihood (MLE) approach to extract the most likely estimate of the gradient with a specific confidence interval. With these tools, quantum supremacy for computing risk may be achievable with a quantum computer whose clock rate is 7 times slower than the pricing itself. This is explained in more detail in Section 7.

[0036] Additionally, this disclosure includes a comparison between quantum gradient estimation algorithms, classical gradient estimation algorithms, and semi-classical gradient estimation algorithms in the context of financial derivatives, summarized in Table 4.

[0037] Additional new contributions are highlighted in the following paragraphs.

[0038] Section 4 describes a numerical study of the quantum gradient estimation algorithms of [2] and [1] for functions of practical interest for financial market risks, and compares the observed oracle costs with theoretical expectations.

[0039] Section 5 describes how to construct a second-order accuracy oracle for quantum gradient estimation, which uses less expensive resources compared to existing methods, for functions computed using quantum amplitude estimation.

[0040] Section 6 describes how to improve the gradient estimation algorithm using classical maximum likelihood estimation (MLE).

[0041] Section 7 provides resource estimates for quantum advantage in financial derivatives pricing.

[0042] Section 9 describes techniques that employ automatic differentiation (AD) techniques on quantum computers that can improve quantum gradient estimation performance in some cases.

[0043] 1.1 Quantum amplitude estimation In various embodiments, the gradient computation algorithm extends a quantum acceleration method for derivatives pricing based on quantum amplitude estimation (QAE), which is further described herein. Let k-dimensional vector x denote a set of market data parameters for a derivative represented as a basis state |x>, and let f(x) denote its price rescaled to satisfy f(x)∈[0,1]. The existence of a unitary operator A in the state

[0044]

number

[0045] where |Ψ0(x)〉 and |Ψ1(x)〉 are arbitrary normalized quantum states. The states |x〉, which represent the market data parameters, act only as inputs to A and can be ignored in the remainder of this section.

[0046] QAE is an operator

[0047]

number

[0048] Estimate f(x) by repeated application of however,

[0049]

number

[0050] and

[0051]

number

[0052] and

[0053]

number

[0054] is the identity matrix. The QAE algorithm is based on the construction of Q such that the state generated after the application of A in equation (1) is an eigenbasis of Q, i.e.

[0055]

number

[0056] where |Ψ + (x)> and |Ψ - (x)> is the eigenvalue e ±2iθ(x) is the eigenvector of Q with f(x)=sin 2 (θ(x)). Let the integer m>0 indicate the number of bits of precision to which you want to estimate the amplitude. m The m-qubit register initialized with

[0057]

number

[0058] and using this register to control different powers of Q, the phase kickback of the eigenvalues ​​of Q is controlled by the control register

[0059]

number

[0060] However,

[0061]

number

[0062] and M=2 m Then, applying the inverse quantum Fourier transform to the first register and measuring it on a computational basis is

[0063]

number

[0064] or

[0065]

number

[0066] This then gives us classically

[0067]

number

[0068] can be used to map the estimate of f(x) to the additional error c p Within

[0069]

number

[0070] To approximate, QAE is a function of M=O(1 / ε p The query complexity of the QAE algorithm is calculated using the oracles A and A † This is measured by counting the number of calls top ) In the case of financial derivatives, the function f is the expected payoff of the derivative contract.

[0071] 2. Gradient Methods for Financial Derivatives Several methods for calculating the slope of financial derivatives can be used. The complexities of these methods are summarized in Table 1.

[0072] 2.1 Greeks

[0073]

number

[0074] Let be a vector of values ​​of d underlying assets at time t.

[0075]

number

[0076] Let the current (today's) market value of these assets be S 0 Let be the paths of the discrete-time multivariate random processes that describe the values ​​of these assets in time, when . Both notations may be used for the paths in the text. The corresponding probability density functions are

[0077]

number

[0078] As shown by:

[0079]

number

[0080] Let be the discounted payoffs of some derivatives on these assets. The prices of the derivatives are

[0081]

number

[0082] where the bar notation has been removed to indicate the switch from a continuous to a discrete model of price. This price gradient is known as the Greeks.

[0083] Example 1 (Delta): Consider a single underlying asset (d=1), and the spot price S 0 The slope of the underlying asset relative to

[0084]

number

[0085] Other commonly used Greeks are the slope of the price with respect to time T (theta), the volatility of the model underlying S (vega), correlations between assets, or other parameters. In general, for some model and / or market parameters x,

[0086]

number

[0087] Calculate and then set

[0088]

number

[0089] Surprisingly, k can be of the order of hundreds or even thousands in practical cases. For this reason, k-scaling improvements represent an advantage in this context.

[0090] 2.2 Classical finite differences

[0091]

number

[0092] Let X be a pricing function with a fixed payoff g. Classically, we use finite differences to find the function,

[0093]

number

[0094] so that expanding f to the first order gives a good approximation of f(x) ≈ f(a) + (xa) ∇ f. The classical scheme of forward finite differences to compute the gradient at a point x0 uses k+1 evaluations of f, one evaluation at x0, and k evaluations displaced from x0 by h in each dimension. This gradient is then

[0095]

number

[0096] is approximated using, where

[0097]

number

[0098] is the i-th normalized basis vector. For a more accurate approximation, a quadratic method using 2k function evaluations is sometimes used, and the gradient is

[0099]

number

[0100] can be approximated by

[0101] Now assume that a function f(x) is evaluated with finite accuracy δ ≥ 0. The associated error of the forward or central finite difference formula is

[0102]

number

[0103] where p=1 is the forward finite difference and p=2 is the central finite difference. Now we look at the asymptotic scaling and say we want to achieve an overall estimation error of ε>0. δ / h+h p ≦ε (9) The step size h can be chosen freely, but improving the accuracy δ is usually associated with increased computational cost. Therefore, the goal is to maximize δ by setting h while still achieving the target accuracy ε. This leads to the following formula: δ = (ε-h p )h=εh-h p+1 (10) Setting the first derivative with respect to h on the right hand side to 0 leads to the following equation: h=(ε / (p+1)) (1 / p) (11) This leads to an optimal δ for a target ε given by: δ = O(ε 1+1 / p ) (12)

[0104] When f(x) is approximated by an algorithm that relies on sampling, the number of samples to target approximation error is δ = (1 / M q ), where M denotes the number of samples and q depends on the convergence rate of the algorithm, q=1 / 2 for classical Monte Carlo simulation and q=1 for QAE. Then combining everything together leads to M=O(ε -(1+1 / p) / q ) (13)

[0105] Thus, with p=2 and q=1 / 2, the complexity of computing k Greeks, computed using central differences with Monte Carlo, is O(k / ε 3 ) This method is sometimes called the classical finite difference (CFD) method.

[0106] 2.3 Finite Differences Using Common Random Numbers Another approach to calculating the Greeks, when derivatives pricing is classically done in Monte Carlo, is to use the second order central difference method in equation (7),

[0107]

number

[0108] and

[0109]

number

[0110] The key to solving this problem is to perform correlated sampling by using the same random numbers in both Monte Carlo evaluations of h. In this way, the statistical fluctuations present in both terms are cancelled out, and the overall error dependence on the discretization step h is effectively removed. In this case, the complexity of evaluating the k Greeks is O(k / ε 2 ), which is also classically optimal. This method is sometimes called CFD-CRN (classical finite difference with common random numbers).

[0111] 2.4 Semiclassical quantum gradients A technique for improving the finite difference method using quantum computing is to use a central difference formula with a quantum amplitude estimation algorithm for each pricing step. This semi-classical quantum gradient (SQG) technique can then be obtained by substituting p=2 and q=1 into equation (13), with a time lag of O(k / ε 1.5) This approach can be improved in a similar way to using CRN for classical finite differences. Using the amplitude estimate

[0112]

number

[0113] and

[0114]

number

[0115] Instead of calculating

[0116]

number

[0117] This means that the quantity

[0118]

number

[0119] This can be done by calculating ω Let denote the payoff for each path ω (see Section 2.1), and then use the amplitude estimation for this quantity. In this case, the output of the amplitude estimation is

[0120]

number

[0121] which gives the expectation value of the finite difference approximation of the i-th derivative. The advantage of this method is that

[0122]

number

[0123] The advantage of this approach is that there is no separate statistical variation associated with each of the gradients individually, and the standard O(1 / ε) scaling of the amplitude estimates is restored. For k gradients, the overall complexity of the method is O(k / ε).

[0124] 2.5 Quantum gradients Some quantum techniques (for computing the Greeks) may use the quantum Fourier transform to compute gradients with improved scaling in dimension k. These quantum algorithms are useful for computing a target function, such as the pricing function in Section 1.1.

[0125]

number

[0126] Includes access to a fractional-phase oracle for point

[0127]

number

[0128] Given and S>0, the operation

[0129]

number

[0130] Execute.

[0131] To estimate the k-dimensional gradient of f at a point x0, the oracle is

[0132]

number

[0133] Assuming that,

[0134]

number

[0135] A k-dimensional hypercube with edge length l around x0 that is sufficiently small such that

[0136]

number

[0137] The measurement is evaluated over a uniform superposition of points δ in , where each dimension is discretized using N points with n = logN qubits. Choosing S = N / l and applying a k-dimensional inverse quantum Fourier transform to the resulting state, the measurement can be made with precision ε = O(1 / N) and high probability.

[0138]

number

[0139] This can be done computationally to obtain an approximation of . The probability depends on the parameter l, but an exemplary "high" probability is a probability greater than 80%. The complexity of these gradient estimation algorithms depends on the resources used to construct the oracle in (15) and the computational complexity used to make the approximation in (16) accurate to within a desired threshold.

[0140]

number

[0141] and the size of x0. As will be explained in more detail in the next section, the computational cost of the gradient estimator depends on the value of S=N / l. Intuitively, the cost of the method increases as the desired accuracy of the gradient estimate increases (larger N) and as the value of l, which indicates the domain over which the function f is evaluated, decreases. The value of N is the input to the gradient estimator that controls the desired accuracy, while the value of l depends on the degree of nonlinearity of the function f in the evaluation domain around x0. The more nonlinear the function f is around x0, the smaller l can be used, so that the function can be approximated by the first-order Taylor expansion used in equation (16).

[0142] 3. Higher-Order Methods for Quantum Gradients To estimate k gradients of a function f with accuracy ε, the oracle in Eq. (15) is

[0143]

number

[0144] times, where D2 is the second derivative of f and an upper bound on the magnitude of N,l defined in Section 2.5. This gives a quadratically improved scaling in k, but it does not improve the error scaling of the semi-classical method. To also improve the scaling in accuracy ε, higher order central difference schemes may be used.

[0145] A 2m-point central difference approximation to the gradient of f at 0 is

[0146]

number

[0147] where the coefficient

[0148]

number

[0149] depends on the choice of m, and for uniform grid spacing it is

[0150]

number

[0151] The phase oracle for the general 2m-point scheme in (17) has the appropriate coefficients

[0152]

number

[0153] It can be constructed by composing individual fractional phase oracles for each of the 2m terms, scaled by: This is a family of algorithms with different phase oracles at different orders m, where S = N / l.

[0154]

number

[0155] For example, the phase oracle of the two-point approximation

[0156]

number

[0157] Oracle

[0158]

number

[0159] and

[0160]

number

[0161] It can be constructed as a product with

[0162] Using these higher-order methods, we can find a class of smooth functions, in particular the Gevrey class G 1 / 2 A function (a function is in Gevrey class G if it satisfies the following formula 20 of Theorem 1) 1 / 2 To estimate the gradient of the function θ(θ) to an accuracy of ε, choose a sufficiently large value for m in the central difference approximation used (in other words, as explained further below).

[0163]

number

[0164] By setting

[0165]

number

[0166] This can be done using the application of a phase oracle to the function. Thus, for this class of smooth functions, gradient estimation using higher order central differences scales quadratically better in the desired accuracy ε. This is stated formally in the following theorem:

[0167] Theorem 1:

[0168]

number

[0169] Let be a fixed constant,

[0170]

number

[0171] is analytical and all

[0172]

number

[0173] and α ∈ [k] j About

[0174]

number

[0175] We assume that the following is satisfied. Using the GAW algorithm,

[0176]

number

[0177] When you set

[0178]

number

[0179] So that the ε-approximation gradient

[0180]

number

[0181] which gives the probability of f or the (fractional) phase oracle

[0182]

number

[0183] has a probability of at least 1-δ using a query of

[0184] Desired

[0185]

number

[0186] In addition to the scaling, the complexity of the GAW algorithm in Theorem 1 includes a factor of logk / δ that arises from extracting the median value after applying a quantum gradient estimation algorithm with probability 1-δ. The following sections are used to establish the governing scaling of the complexity with respect to k.

[0187]

number

[0188] While we focus on the log(k / δ) factor, Section 6 shows how the log(k / δ) factor can be eliminated by using classical maximum likelihood estimation (MLE). This approach not only reduces the overall computational complexity of the algorithm, but also provides a concrete confidence interval for a given confidence level.

[0189] Table 1 summarizes the scaling in k and c of the algorithms described in this section.

[0190] [Table 1]

[0191] 3.1 Generating Phase Oracles from Probability Oracles Since the function over which the gradient is computed is accessible via a probability oracle of the form of equation (1), the corresponding phase oracle in equation (15) can be used. To implement the quantum gradient methods described in this section using phase oracles, block coding techniques can be used.

[0192] Definition 1: Suppose A is an operator on a Hilbert space H, then

[0193]

number

[0194] The unitary U acting on is

[0195]

number

[0196] It can be said that this is a block coding of A in the case of

[0197] Intuitively, a block coding U is a unitary whose top-left block contains A.

number

[0198] For the probability oracle A in equation (1),

[0199]

number

[0200] And from Definition 1,

[0201]

number

[0202] is a block coding of a diagonal matrix H with diagonal entries (1-2f(x)). With access to this block coding, Hamiltonian simulation methods (e.g., Non-Patent Document 4 or Non-Patent Document 5) can be implemented in O(|t|+ln(1 / ε phase )) for U and U † By applying the unitary eitH ε phase Thus, with block coding of f(x) and t = 2πS, the Hamiltonian simulation can be done by deriving the phase oracle from (15):

[0203]

number

[0204] ε phase (As alluded to earlier, ε approximation of a quantity means that the approximation is within error ε of the quantity.) Note that the block coding of equation (23) allows the creation of a phase oracle with a phase of 1-2f(x) instead of f(x) of equation (15), but the first term becomes a global phase that can be ignored, and the factor of -2 can be absorbed into the factor t of the Hamiltonian simulation.

[0205] Hamiltonian simulation can be used to realize unitary evolution, but some of these methods (such as, for example, [4] and [5]) may require post-selection to be considered for end-to-end resource estimation. On the other hand, [6] does not require post-selection and instead succeeds with arbitrarily high probability 1-δ and has O(|t|+ln(1 / ε phase We introduce a coherent Hamiltonian simulation method that scales as ε(1 / δ)+ln(1 / δ) for the target t and the approximation error ε phase and δ=2ε phase For , the coherent Hamiltonian simulation algorithm computes U and its inverse a total of

[0206]

number

[0207] where (1) β ∈ (0,1) is a parameter selected by the user, and (2) r(T, ε) = |T|e W(ln(1 / ε) / |T|) where W(x) is the Lambert W function, (3)

[0208]

number

[0209] It is.

[0210] Using this method, the m-th phase oracle ε phase The total number of oracle (equation (1)) calls used to construct the approximation is

[0211]

number

[0212] is given by:

[0213] The optimal value of β is the target approximation error ε phase In the cases studied in this disclosure, β=0.5 is an improved (e.g., optimal) choice and can be used in equation (25).

[0214] 4. Resource Estimation for Quantum Gradient Methods This section describes the asymptotic resource estimation for the gradient methods described previously. Representative parameters from the financial domain are Greeks with k=1000 and ε=10 -3 For the GAW resource estimation, the smoothing condition in Theorem 1 with smoothing parameter c = 1 is applied to the problem at hand, and the Hamiltonian simulation phase error is ε phase =10 -4 where c=1 is used to estimate the possible usefulness of the algorithm in the best case scenario of smoothing from Theorem 1. The phase error εphase can be treated as a free parameter at this point, but 10 -4 The particular choice of is motivated by numerical simulations which show that this is a good choice. The numerical simulations and the effects of phase errors are described in more detail in subsequent sections. The finite difference approximation is

[0215]

number

[0216] and the interval parameter is set to

[0217]

number

[0218] Using a proof from [2], the number of oracle invocations to achieve the target error can be estimated. In these example estimates, only asymptotic scaling is considered (assuming the remaining constants are a factor of 1), and ε phase Note that is assumed to be small enough not to affect performance. This is 6.3×10 7 This results in oracle invocations.

[0219] This can be compared asymptotically with the performance of both classical finite differences (with or without common random numbers) and semiclassical quantum finite differences. These results are summarized in Table 2. In one aspect, the optimal value of the discretization step h calculated in Section 2.2 is chosen to minimize the gradient estimation error ε, with ε=10−3 chosen for these benchmarks.

[0220] Table 2 shows, using parameters from the proof in [2], that the GAW algorithm fails to provide an advantage compared to semi-classical methods for k as high as 103. However, this analysis gives an indication of how the different methods compare theoretically, although the estimates are based on asymptotic bounds with parameters that may be loose in practice and depend on the smoothness of the functions considered. The following sections numerically analyze the performance of the GAW method on examples that represent some practical cases in finance.

[0221] [Table 2]

[0222] 4.1GAW Numerical Estimates The GAW method described in Section 3 can be applied when the target function satisfies the conditions of Theorem 1.

[0223]

number

[0224] We provide a quantum algorithm for gradient estimation that scales with . However, for most relevant financial models of interest, closed-form solutions are not available that can be examined to check whether they satisfy the required smoothness condition of the theorem (e.g., numerical methods are used to approximate functions of interest that do not have closed-form solutions). A closed-form representation, as used herein, is a mathematical representation that uses a finite number of standard operations. It may include constants, variables, operations (e.g., +-×÷), and functions (e.g., n-th roots, exponentials, logarithms, trigonometric functions, and inverse hyperbolic functions), but does not include limits, differentiation, or integration. In some embodiments, if a numerical method is used (e.g., must be used) to approximate a function, the function does not have a closed-form solution.

[0225] Therefore, this section describes the numerical investigation of the behavior of higher order methods of the GAW gradient estimation algorithm for two financial use cases: (a) simple (vanilla) European call options, for which there are available analytical closed-form solutions that allow benchmarking of the algorithm's performance for accurate gradients, and (b) path-dependent basket options with knock-in features that do not have a known analytical solution and are classically evaluated in practice using Monte Carlo methods. In particular, this section examines the central difference approximation order m and interval l used for sufficiently accurate gradient estimation and compares it with the theoretical values ​​in Equation (26) and Equation (27), respectively. This section focuses on these two parameters, since they determine the overall complexity of the algorithm, which can then be compared to other methods in Table 2.

[0226] Since the resulting quantum circuits are prohibitively wide and deep for practical numerical simulation, the following practical method is adopted to simulate the performance of the algorithm. To estimate the k Greeks using n-bit precision, we use the amplitude of (18) classically calculated for the order m of the selected derivative to obtain k*2 n Initialize an array of dimension k*2 of the array. n A classical inverse Fourier transform of dimension 1 is performed, resulting in a probability distribution that is the output of the GAW algorithm before measurements. The phase error ε from the Hamiltonian simulation phase To consider >0, the interval [-ε phase ,ε phase ] is added to each encoded phase in equation (18).

[0227] 4.1.1 Vanilla options A simple example in derivatives pricing is the European call option, whose price depends on the performance of a single asset at a given future time (expiry date) against a reference price (the strike). In the Black-Scholes-Merton model, where the asset undergoes Geometric Brownian Motion (GBM), a call option on a non-dividend paying asset is

[0228]

number

[0229] We have a closed-form solution given by where S is the asset price today, K is the strike of the option, r is the risk-free rate of return, σ is the annualized volatility of the asset, T is the time until the option expires, and N(x) denotes the CDF of the standard normal distribution. A quantum algorithm can be tested for the four Greeks of this option / model, namely delta = ∂C / ∂S, rho = ∂C / ∂r, vega = ∂C / ∂σ, and theta = ∂C / ∂T. Numerical simulations of the GAW algorithm can be performed to increase k (the number of Greeks to compute simultaneously) and the central difference approximation order m ∈ [1,4]. In each case, the search is performed for the maximum value of the interval l for which the algorithm produces estimates that are ε close to the correct value with a probability of 85% or higher for each Greek. ε = 2 × 10 using n = [log(1 / ε)] = 6 qubits in each dimension. -2 (Note that these numerical simulations scale exponentially in this parameter, since we are modeling quantum circuits on these qubits.)

[0230] For each value of k, we find the total number of oracle calls N from (25) while maintaining a success probability of at least 85%. oThe optimum values ​​of (m,l) found numerically that minimize ε are compared to those used in the proof of Theorem 1 given by (26) and (27), assuming the smoothest possible parameter c=1. The approximation error from the Hamiltonian simulation is phase =10 -4 and is included as a source of error in the numerical simulations. The gradient of the vanilla option in equation (28) is evaluated at points (S, r, σ, T) = (99.5, 1%, 20%, 0.1) with K = 100, chosen to explore a domain where the parameter values ​​are reasonably realistic from a financial point of view, but at the same time the function is as nonlinear as possible (simple finite difference methods already work well when the function is (approximately) linear). Table 3 shows the results of the numerical simulations and the corresponding theoretical estimates from Theorem 1, and Figure 1 shows the probability distributions obtained from the quantum gradient estimation algorithm for each Greek for the case k = 4 in Table 3. To simulate the algorithm for increasing values ​​of k, we choose k of the four parameters S, r, σ, T for which the gradient is calculated, and fix the values ​​of the remaining 4-k parameters. For the case k ∈ [2, 3], the choice of which parameters to fix, the simulation results in Table 3 are qualitatively independent of the choice.

[0231] Figure 1 shows the probability distributions resulting from numerical simulations of the GAW quantum gradient estimation algorithm with m=3 and l=0.58 for the four Greeks (delta, rho, vega, theta) of vanilla options in equation (28), along with the corresponding exact values ​​shown as dashed vertical lines. Measurements of the registers corresponding to each Greek deviate by up to ε=1 / N=2×10 from the exact value with a probability of more than 85%. -2 This results in a value that is far away.

[0232] [Table 3]

[0233] Interestingly, Table 3 shows that the query complexity used to estimate gradients for vanilla options with high probability in practice is several orders of magnitude smaller than would be expected from the parameters used in the proof of Theorem 1. Since the types of vanilla options explored in this section are primarily motivating examples of relevance to finance that are easier to analyze, the following sections focus on more complex derivatives for which gradient estimation is actually used for business purposes.

[0234] 4.1.2 Path-dependent basket options The previous section described a numerical investigation of the performance of the GAW algorithm for vanilla options for which there is a known analytical solution that can be used to benchmark the performance of the algorithm against the exact gradient of the model. A similar analysis may be performed for multi-underlying path-dependent options that do not have a known closed-form solution, with the query complexity from the numerical simulations compared to the theoretical complexity from Theorem 1 and to the complexity of the SQG method (Section 2.4), which computes gradients using finite differences using values ​​estimated using QAE. As in the previous section, the approximation error from the Hamiltonian simulations is ε=2×10 -2 may be set to

[0235] The options considered in this section have volatilities σ1=20%, σ2=20%, σ3=10%, and spot prices

[0236]

number

[0237] The risk-free rate is set to r=1% and the option expires in T=3 years. Weights

[0238]

number

[0239] The asset price

[0240]

number

[0241] The weighted sum of is the t B = [T / 5*i] and the option payoff is

[0242]

number

[0243] where the strikes are set to K=1.0 and B=2.5. The option is said to have a knock-in feature since it will only make profits (in other words, be knocked in) if the weighted sum of the observed options exceeds the knock-in barrier B at any of the pre-determined times before the expiry of the contract. The GAW algorithm is simulated to estimate the four gradients of the price V of this option contract, namely the three deltas (∂V / ∂Si) and the vega ∂V / ∂σ1 for the first asset. Since this option does not have a closed-form solution, the option is priced using classical Monte Carlo with 106 paths and finite differences to calculate the expected gradients, which can be used as a benchmark for the quantum algorithm. The quantum algorithm may be simulated using n=4 qubits for each gradient register, which limits the target error of the algorithm to ε≦1 / 2 4= 0.0625. Using m = 1 and l = 0.25, the simulated GAW algorithm gives estimates close to ε with 85% probability for each Greek. The SQG method uses a 1 / ε call to the Q operator of the QAE, where each Q contains two calls to the unitary A in equation (1) and its inverse, so the complexity for k Greeks is 2k / ε. Because the SQG method calculates the payoff twice to construct the finite differences (equation (14)), the A operator is about twice as large as a normal pricing oracle. To more accurately compare the query complexity to other quantum methods, a factor of two complexity is sometimes included, giving a total of 4k / ε.

[0244] Table 4 shows the query complexities and parameters from numerical simulations of the GAW method for this path-dependent basket option, along with (a) the asymptotic estimates from Theorem 1, (b) the query complexities of the SQG method, and (c) the query complexities of the CFD and CFD-CRN methods, all for the same target approximation error. For the CFD and CFD-CRN methods, the reported query complexities are the total number of Monte Carlo paths used to evaluate the k Greeks within a numerically computed target approximation error ε with probability 85%. The table also includes the parameters and resources used for the same calculation using the Simulation-Free Quantum Gradient (SFQG) algorithm described in the next section. The resulting probability distributions for all the Greeks are shown in Figure 2 along with the MC-estimated "true" values.

[0245] Figure 2 shows the probability distributions resulting from numerical simulations of the GAW quantum gradient estimation algorithm with m=1 and l=0.25 for the four Greeks of the path-dependent basket option in Section 4.1.2, over a 10 6 The measurements are shown together with the values ​​estimated using classical Monte Carlo (MC) with a path of ε (vertical dashed line). The measurements provide estimates for each Greek with an 85% or higher probability of having an error ε≦0.0625.

[0246] 5. Simulation-Free Quantum Gradient (SFQG) Method This section describes the construction of a second-order exact quantum gradient algorithm, corresponding to m=1 in (17). This is of higher order than the algorithm in [1] (which has unfavorable complexity scaling), but

[0247]

number

[0248] This is a lower order than that required from the analysis in [2] to guarantee the scaling of . The advantage is that for derivatives pricing, it provides an explicit construction of phase oracles without using block coding or Hamiltonian simulation. This is sometimes called the Simulation-Free Quantum Gradients (SFQG) approach.

[0249] To illustrate this configuration, this section first compares the derivatives pricing in Section 1.1 with the first phase oracle.

[0250]

number

[0251] Then, as defined in equation (18),

[0252]

number

[0253] We show how to construct a quadratic expansion to construct the GAW method. This algorithm can be simulated to estimate the Greeks of the basket options from Section 4.1.2, allowing for a comparison of performance with the corresponding second-order exact GAW method.

[0254] 5.1 First Pricing Phase Oracle To apply the quantum gradient algorithm to estimate the gradient of a function f(x), a phase oracle of the type of Equation (15) can be constructed. This oracle is a k-dimensional hypercube G k =[-l / 2,l / 2] k The A operator in equation (1) can be evaluated for a superposition of points x in x, where each dimension is discretized using N points with n=logN qubits, and l is chosen small enough so that f(x) is approximately linear in its domain (so that the approximation in equation 16 holds). For simplicity, let x0=0, since gradient estimation at other points can be achieved by redefining the function f. Evaluating the A operator in equation (1) for all points x in the superposition is

[0255]

number

[0256] However, as in equation (3),

[0257]

number

[0258] Next, the Grover operator

[0259]

number

[0260] Define, where:

[0261]

number

[0262] and

[0263]

number

[0264] For all i∈[1,k], |(∂θ(x) / ∂x i )| is bounded by D, S=N / Dl, and applying the Q operator to the state of equation (30) πS times gives

[0265]

number

[0266] Since the states |ψ±(x)> in (31) depend on x, care is usually taken to ensure that they do not interfere with the phase kickback to the |x> register. As we will explain in Section 9, the structure of the A operator used in derivatives pricing allows the phase kickback to occur correctly.

[0267]

number

[0268] Note that application of is close to the desired phase oracle for quantum gradient estimation, but (i) for θ instead of f, and (ii) is superimposed across the sought positive and negative phases.

[0269] Inverse quantum Fourier transform for the first register

[0270]

number

[0271] gives an estimate of the derivative of θ(x) at x0 as follows:

[0272]

number

[0273] This method gives us an oracle for θ, but not for f directly. However, θ and f are related by the fact that f(x)=sin 2 Since we are related via (θ(x)), the derivative of θ can be calculated and the chain rule applied to get the gradient of f.

[0274]

number

[0275] This uses knowledge of θ(x0), which can be calculated through the QAE via the operator in equation (1).

[0276] By adding a dummy dimension to f with a gradient of known sign, it is possible to distinguish between the positive and negative cases. For example, the function

[0277]

number

[0278] We can transform it by checking the sign of the k+1 derivative to see if it is positive or negative.

[0279] In this setting, the value of πS = πN / Dl must be representable as an integer, since it indicates the number of times the Q operator in equation (31) is called. n Since is an integer, this can be achieved by choosing a value for D such that π / Dl is also an integer. After application of the Q operator, the measured gradient of equation (32) is classically multiplied an integer number of times by the choice of D to recover the magnitude of the gradient.

[0280] 5.2 Secondary Pricing Phase Oracle A possible higher-order extension is to construct an oracle that encodes the two-point approximation shown in (19), which corresponds to m=1 in (17). In the QAE setting,

[0281]

number

[0282] Since the gradient of is computed instead of f(x), the constructed oracle is

[0283]

number

[0284] However,

[0285]

number

[0286] In general, higher-order oracles of this form can be constructed as a product of two separate oracles with opposite phases. This idea can be used to construct the oracle of equation (34) using appropriately defined oracles. A + Let us denote by A operator in equation (1) and let A be the operator that acts as a probability oracle for the value f(-x). -

[0287]

number

[0288] and the corresponding Grover operator

[0289]

number

[0290] and

[0291]

number

[0292] From equation (31), the oracle

[0293]

number

[0294] and

[0295]

number

[0296] The product of this is the state

[0297]

number

[0298] where ψ i Q + and Q - and can be neglected for the remainder of the algorithm. The first two terms contain the appropriate phase kickback (including the sign) for the two-point approximation of (34) with 50% joint probability, while the last two terms are θ(x)+θ(-x)=θ(x0)+O(∂ 2 θ(x) / ∂x 2), which produces a probability peak near 0 instead of the gradient, with a similar joint 50% probability. By adding a dummy variable to the function as explained at the end of section 5.1, it is possible to distinguish which eigenstate we are in by measuring the gradient of the dummy variable after applying the inverse quantum Fourier transform. Since the gradient with respect to the dummy variable is known by construction, measuring a positive (negative) gradient of the dummy variable means measuring a positive (negative) gradient with respect to the other variable. Otherwise, if a zero is measured in the dummy variable register, the measurement can be ignored. This means that additional post-processing is used for this method, where 50% of the shots are discarded. The total number of shots used depends on the desired accuracy of the estimation, and this choice is explained in more detail in section 6. A schematic diagram of the SFQG method is shown in Figure 3. To generate the correct phase of equation (36), the operator Q + and Q - Note that, applies to separate registers.

[0299] Figure 3 shows the quantum circuit diagram of the SFQG method used to generate the state of equation (36). First, the superposition of points Σ x To generate |x>, a Hadamard gate is applied to k registers of n qubits each. A CNOT gate (controlled by qubits in a first set of k registers of n qubits) is then used to create a copy of each |x> (in a second set of k registers of n qubits), which is then used to generate the oracle at the expense of an extra qubit.

[0300]

number

[0301] and

[0302]

number

[0303] After uncomputing a copy using another set of CNOT gates (controlled by qubits in a first set of k registers of n qubits), an n-dimensional inverse quantum Fourier transform is applied to each of the first k registers in the first set to give the state of equation (36), which is then measured to estimate the k gradients.

[0304] Using the operators to construct the state of equation (36), a numerical simulation can be used for the SFQG algorithm to compute the Greeks of the basket options of Section 4.1.2. In this case, the algorithm is used to compute the gradients ∂θ / ∂S1, ∂θ / ∂S2, ∂θ / ∂S3, ∂θ / ∂σ1, where

[0305]

number

[0306] where V is the option price, from which the Greeks of the option can be estimated using (33), assuming that the contract has already been priced. As in the simulation of the GAW algorithm in section 4.1.2, a search is performed for parameter values ​​l that estimate a gradient within ε≦0.0625 with a success probability of 85% or more, using n=4 qubits for each gradient register. Here, the probability of obtaining the discarded state in (36) in post-processing can be ignored. Intuitively, given a 50% failure probability, one can consider the effective oracle cost to be twice that calculated from this simulation, but as shown in section 6, there are more efficient ways of combining measurements to obtain gradient estimates and confidence intervals. The complexity of the algorithm in terms of the number of successive calls to the A operator in (1) used to construct the state in (36) is πS=πN / l, where N=2 n It is.

[0307] The probability distributions in the non-discarded states for all four gradients after application of the SFQG algorithm are shown in Figure 4, and the parameter values ​​and query complexities of the algorithm are shown in Table 4 along with the other quantum, classical, and semiclassical methods studied in this disclosure. The complexity of the classical method is estimated by computing the Greeks by finite differences using 3000 Monte Carlo simulations of basket option prices and exploring the number of Monte Carlo paths that give the same target error ε≦0.0625 as the quantum and semiclassical methods with a probability of 85% or higher.

[0308] Figure 4 shows the probability distributions resulting from a numerical simulation of the second-order exact SFQG algorithm with m=1 and l=0.25 for the four Greeks of the path-dependent basket option of Section 4.1.2, together with the exact values ​​estimated using classical Monte Carlo simulations (vertical dashed lines). The algorithm is based on the function

[0309]

number

[0310] where V is the option price, from which the option's Greeks can be estimated using equation (33). As in Figure 2, the measurements give estimates of each Greek within error ε≦0.0625 with a probability of 85% or higher.

[0311] [Table 4]

[0312] 6. Quantum Gradient Estimation Using Maximum Likelihood Estimation The success of the quantum gradient estimation algorithms studied in this disclosure (e.g., the SFQG method), as well as the algorithms of [1] and [2], can be judged by their ability to estimate the gradient within ε with high probability. In [2], the authors propose to repeat the gradient estimation algorithm O(log(k / ρ)) times for k gradients and take the median to obtain an estimate within error ε with a probability of at least (1-ρ). However, this approach has two main drawbacks: a) it does not characterize the correctness of the output with a precise confidence interval, and b) the output is a discretized hypercube of Eq. (16).

[0313]

number

[0314] These issues may apply to amplitude estimation algorithms as well.

[0315] Maximum likelihood estimation (MLE) can address these issues for amplitude estimation algorithms. For example, the quantum circuit is sampled two or more times and the results are classically post-processed with MLE. This section shows that MLE methods can be applied incrementally to the quantum gradient estimation algorithm to improve the final estimate.

[0316] A probability distribution p with unknown parameters g and sampled data x i Given (i=1,...,M), the MLE is the estimate of g.

[0317]

number

[0318] This is done by maximizing the log-likelihood logL, where

[0319]

number

[0320] and g' is a candidate to be evaluated.

[0321]

number

[0322] This means that if g' is a true parameter, then the data x i The general quantum gradient estimation algorithm from section 2.5 uses the region l where the function whose gradient it is computing is approximately linear. In this case, the probability distribution after applying the inverse quantum Fourier transform to equation (16) is the same as that for phase estimation,

[0323]

number

[0324] where Δ=(x-g') and N is the number of possible measurements.

[0325] Estimating the MLE

[0326]

number

[0327] The confidence interval for can be derived using the likelihood ratio (LR). The confidence interval at the (1-α) confidence level is

[0328]

number

[0329] where g' is a value that satisfies:

[0330]

number

[0331] is χ 2 The (1-α) quantile of the distribution is shown.

[0332] Figure 5a shows the discrete probability distribution before the Vega ∂θ / ∂σ measures (bars) for the basket options from Figure 4 are fitted to the theoretical distribution p(x) (solid line) of Equation 38. The maximum of the distribution occurs at the exact analytical value (vertical dashed line).

[0333] Figure 5b shows the global maximum of the log-likelihood logL(g') (dots) that provides a better estimate of the true value (vertical dashed line) than the most likely outcome of sampling from the probability distribution and taking the median (triangles). The log-likelihood plot was generated by sampling the quantum gradient estimator 30 times and shows that ε ≤ 2 × 10 with a confidence level of 1-α = 0.68. -3 It allows us to estimate this Greek within 100%. When using the SFQG method, the probability of discarding a measurement is 50%, and therefore the expected number of shots for this target accuracy is 60.

[0334] Figures 5a and 5b show how MLE can be used to improve the Vega estimate (∂θ / ∂σ) for the basket option from Figure 4 calculated using the Simulation-Free Quantum Gradient (SFQG) method. As explained in Section 5.2, the SFQG method produces a 50% probability peak around the gradient, including the sign, and another 50% peak around 0, which it ignores. The technique can distinguish between these cases by adding a dummy variable to the function with known gradient, so that only the cases where the gradient is measured may be considered, and the number of shots may be doubled to account for the discarded measurements. Figure 5a shows that the final probability distribution after application of the quantum gradient estimation algorithm (labeled "quantum gradient") does indeed fit Equation (38) (plotted in Figure 5a and labeled "p(x)"), and therefore the QAE with MLE results can also be used here. For clarity, the measurement results are plotted only in the interval that contains most of the probability mass.

[0335] Figure 5b plots the log-likelihood logL(g') as a function of g' over the same interval when the quantum gradient circuit generating the probability distribution in Figure 5a is sampled 30 times. In Figure 5a, the MLE estimate

[0336]

number

[0337] (dot) is the global maximum of the function (labeled "LogL(x)") that is very close to the true value g (vertical dotted line) and significantly better than the median of the final probability distribution (triangle). In practice, the SFQG method may use twice the number of shots (60) to account for the shots discarded due to the extra term in equation (36).

[0338] One important advantage of the MLE approach is that because the final estimate is not constrained to be one of N possible discrete values, the value of N can be reduced at the expense of increasing the number of samples taken, thus lowering the overall width and depth of the quantum circuit. For example, the distance between the N possible discrete values ​​in Figure 5a is 0.0625, while the error ε in the MLE estimate in Figure 5b is ε = 4 × 10 -3 It is.

[0339] Although MLE post-processing uses additional classical computational cost, computing the MLE in this setting is done by growing (e.g., maximizing) a concave function over a one-dimensional compact interval, which can be ignored in the overall complexity analysis.

[0340] 7. Updated estimates of quantum supremacy The resource estimates in [3] suggest that quantum superiority for derivatives pricing over classical Monte Carlo methods is approximately 10 10 A code distance of approximately 10 per second can support logical operations of 7 In particular, we have established that a quantum processor capable of executing T gates of ε≦2×10 is required. In particular, we have used the reparameterization technique introduced in [3] to reconstruct the autoinvokable contracts studied in [3] with ε≦2×10 -3 Pricing with confidence level 1-α=0.68 within 5×10 7 Therefore, to match the classical Monte Carlo pricing time estimated to be 1 second, the logical quantum clock may use a rate of 50 MHz. The same reparameterization technique from [3] may be used to construct the oracle size estimates and estimate the T-depth of the quantum circuit for the SFQG method, computing the four Greeks of basket options in Section 4.1.2 and Table 4. The same ε≦2×10 -3The SFQG method can be used with the parameters of Figure 4 and the maximum likelihood estimation (MLE) technique from Section 6 using 60 shots to compute the Greeks within a confidence level of 1-α=0.68 (see Figure 5b). Using these parameters, it can be seen that if an end-to-end circuit uses 12k logical qubits and performs T-gates with a code distance capable of supporting 10 logical operations, then the total T-depth of the SFQG circuit multiplied by the number of shots is 5.5×10 7 This gives an estimate that the Greeks of a basket option can be classically estimated in 8 seconds (4 Greeks, 2 pricings per Greek), assuming that the contract can be classically priced in 1 second using Monte Carlo, within the same error and confidence level, and that the Greeks can be calculated using second-order finite difference methods applied to Monte Carlo pricing. Thus, quantum advantage using the SFQG method in calculating the Greeks of this derivatives contract may involve executing T-gates at a rate of 7 MHz, which is 7 times lower than the estimate in [3] for quantum advantage in derivatives pricing. Furthermore, since the quantum gradient circuit is sampled 60 times when the MLE approach is used (see FIG. 5), providing parallel access to 60 QPUs on which the quantum circuits can be loaded and sampled simultaneously would allow achieving the same execution time as a serial run, if the logical clock rate of each device is around 100 kHz, which is closer to the current estimate of the achievable logical clock rate of around 10 kHz.

[0341] Although the computational costs for numerical simulation of the SFQG method currently limit the analysis to Greeks with a maximum of k=4 for practically relevant use cases, the SFQG method can be favorably scaled to higher dimensional derivatives pricing problems, motivated by the fact that most, if not all, derivative contracts of practical interest have piecewise linear payoffs. Theorem 1 shows that for the class of functions considered,

[0342]

number

[0343] Although the SFQG method offers a complexity of O(log(k) / ε) for n>1, m∈[1,n-1], it does not preclude speedups of higher orders in k for smoother functions. For example, the SFQG method applied to simple polynomial functions can achieve a complexity of O(log(k) / ε). Derivatives contracts are often written as

[0344]

number

[0345] For example, a basket option in Section 4.1.2 that does not have a knock-in feature.

[0346]

number

[0347] The absence of such higher order terms makes it possible to choose values ​​of m and l that result in an overall oracle cost smaller than that specified by Theorem 1 in the application of the SFQG method. An illustrative upper bound for the potential advantage over classical Monte Carlo is when the function is at most a quadratic polynomial in all k variables, resulting in an overall speedup of O(k). Thus, the possible degree of quantum speedup of the SFQG method may depend on the nature of the price function of each derivative. Assuming that the SFQG method can be applied to larger numbers of k (as discussed above), one or more of the above advantages may also apply to the GAW method with m=1. However, as shown by Table 4, the SFQG method is even cheaper to construct than the GAW method.

[0348] 8. Phase kickback In the simulation-free quantum gradient (SFQG) approach described in Section 5, care can be taken to ensure that the proper phase kickback occurs in (31) (e.g., to obtain the correct gradient after application of the inverse quantum Fourier transform). Equation (31) generally introduces a state |Ψ that interferes with the subsequent inverse quantum Fourier transform through its dependence on x. ± (x)>. The application of the Q operator in the SFQG method is

[0349]

number

[0350] Applying the inverse quantum Fourier transform to the |x> register produces

[0351]

number

[0352] The probability of measuring the value |z> in the first register is given by

[0353]

number

[0354] In the derivatives pricing context considered in this disclosure, the A operator may implement a reparameterization method (e.g., the reparameterization method of [3]). To do this, the A operator can be written as a product of two operators G and F. The G operator determines the number and pricing of assets in the derivatives contract.

[0355]

number

[0356] Load the standard normal distribution corresponding to the time step used ini is independent of any market parameters. The F operator then computes the payoff g(i) using quantum operations for |i>, which is then calculated as the ancilla qubit

[0357]

number

[0358] is rotated to an amplitude of

[0359] After the last rotation, the register |g(i)> can be uncomputed, so the operator

[0360]

number

[0361] The overall effect of

[0362]

number

[0363] where g(i)∈[0,1]. When A is evaluated in a superposition on the registers |x>, which represent fine adjustments to the input market parameters, as shown in (30), the resulting state is

[0364]

number

[0365] However,

[0366]

number

[0367] and

[0368]

number

[0369] where k is the dimension of the gradient we are estimating and N=2 n and n qubits are superpositions in each dimension Σ x |x>. Choosing x≪1 means that the function θ(x) is approximately linear in the neighborhood of x=0. In this regime, |Ψ in (46) + x> state is <Ψ + (x')|Ψ - (x)>=<Ψ - (x')|Ψ + (x)>≒0 and <Ψ + (x')|Ψ + (x)>=<Ψ - (x')|Ψ - We give (x)>≈1. Then the probability of measuring the value |z> in equation (41) is

[0370]

number

[0371] which is similar to standard quantum phase estimation. Note that since the x dependence in (46) is limited to the amplitude of the last qubit, the application of the Q operator in (39) induces the correct phase kickback to the |x> register. With probability p i If x becomes dependent on x, or if other qubit registers remain entangled with x, then the phase kickback will fail.

[0372] FIG. 6 shows the simulated measurement results after creation of the state of (39) and subsequent inverse quantum Fourier transformation for an example when the A operator is of the form of (45) and k=1. i is Σ i p i= 1, and g(i,x) = sin at b = 0.405 2 It is obtained from a normal distribution with unit variance defined for three qubits (m=3) normalized to be (b*i+2x). After applying the quantum Fourier transform and measuring the |x> register, we can obtain an estimate of the gradient dθ / dx (or -dθ / dx as described by equation (32)) at x=0, where

[0373]

number

[0374] and a = Σ p p i sin 2 (b*i+2x). For the chosen parameter values, dθ / dx≈0.253 at x=0. Three qubits may be used for the register |x>(n=3), which allows for resolution of the gradient with an accuracy of 1 / 8=0.125. The value of b was chosen for clarity such that the resulting gradient is close to one of the values ​​that can be accurately represented with three qubits. One might choose l=π / 256 and D=1, so that the Q operator can be applied S=N / Dl=2048 times. The A operator used in this simulation is of the form (45), so that the correct phase is kicked back into the |×> register, and the inverse quantum Fourier transform produces two probability peaks around ±dθ / dx, where dθ / dx≈0.253.

[0375] 6 shows a simulated measurement result after application of equation (31) and subsequent inverse quantum Fourier transformation when the A operator is of the form of equation (45). The measurement result corresponds to an estimate of the gradient dθ / dx at x=0, where

[0376]

number

[0377] and a=Σ at b=0.405.i p i sin 2 (b*i+2x), with p=[0.03149738, 0.08388914, 0.16118575, 0.22342773, 0.22342773, 0.16118575, 0.08388914, 0.03149738]. The probabilities peak near ±0.25, which is the closest value that can be represented with n=3 qubits for the expected dθ / dx≈0.253.

[0378] 9. Automatic Differentiation and Multi-Objective QAE Another way to compute the gradient (e.g. in the considered setting) is automatic differentiation (AD). AD applies the chain rule iteratively to all the basic arithmetic operations used to compute the objective function, tracking the analytical gradient throughout the computation. In many practical applications, the gradient can be computed with a constant overhead, independent of the dimension. In some cases, it can be said that the overhead is reduced by a factor of four compared to evaluating only the function, at the expense of larger memory requirements.

[0379] The expected value for a payoff function g, a random variable S, and some given parameters x

[0380]

number

[0381] and the corresponding gradient

[0382]

number

[0383] Suppose we want to estimate

[0384] To construct the probability oracle A(x) used by the QAE for a fixed x, we first create a weighted superposition of (e.g., all) scenarios, then evaluate the corresponding payoff for each scenario, and determine whether the probability of measuring |1> in the final qubit is

[0385]

number

[0386] To correspond to

[0387]

number

[0388] We can prepare a target qubit to obtain s j Let denote the possible realizations of S represented by m qubits, and p j denotes the corresponding probability.

[0389] All scenarios |s j >, the payoff |g(s j Quantum operations can be applied to compute s , x . Thus, for each s j For AD, the gradient ∇ x g(s j ,x) can be used in the same way as the classical one, but using a reversible implementation uses at most twice the resources used classically. Thus, with a constant overhead compared to evaluating the expectation, this results in the following state:

[0390]

number

[0391] The following shows how to use QAE to retrieve multiple objectives defined with the same random variables, which is then immediately applied to the gradient constructed in equation (50).

[0392] Let S be a random variable and the realization of S be

[0393]

number

[0394] A set of functions f that map to i Assume that i = 1,..,k. Furthermore, for all i, the expected value

[0395]

number

[0396] We are interested in estimating

[0397]

number

[0398] Suppose we can construct a state of the form

[0399] value

[0400]

number

[0401] To estimate, first, we define i k additional m-qubit registers initialized with |c i > and secondly, to another register.

[0402]

number

[0403] Use quantum computing to calculate. In other words,

[0404]

number

[0405] It is possible to construct an operator that acts as a function by adding a target qubit and applying a rotation controlled by the last register, which also yields the function

[0406]

number

[0407] It can be used to construct a probability oracle A(c) corresponding to k, which is a linear function in c, and using the quantum gradient algorithm with respect to c is the readout of all k expectations.

[0408]

number

[0409] results.

[0410] f i The value of is the weighted sum of Eq. (52)

[0411]

number

[0412] If the construction function is linear in c, then the resulting complexity of the quantum gradient algorithm for a target accuracy ε>0 scales as O(1 / ε) independent of k (ignoring the logarithmic term). A general-objective QAE can be computed with logarithmic depth by using k m additional qubits, as well as the weighted sum in equation (52), using a divide-and-conquer summation scheme. On the other hand, if equation (52) is normalized by dividing the weighted sum by a factor D, the complexity of the algorithm becomes

[0413]

number

[0414] For all i, f i When .about.1, we may choose D.about.1 / k, which adds a factor of k back into the complexity of the algorithm, negating the advantages of this approach.

[0415] Since equation (50) has the desired (e.g., required) shape, a multi-objective QAE can be applied to evaluate the gradient evaluated using the AD implemented with quantum operations. If no additional regularization is desired (e.g., not required), then the AD and (multi-objective) QAE can be applied with a dimension-independent running time.

[0416]

number

[0417] and can be combined to obtain a gradient algorithm with quadratic speedup in accuracy. Thus, like the classical ones, AD is a promising approach for estimating market risk, with significant advantages over finite difference methods.

[0418] 10. Consideration In particular, this disclosure provides new applications of and new modifications to the GAW algorithm. In particular, this disclosure describes the use of a modified version of the GAW algorithm (m=1) to compute the gradients (greeks) of financial derivatives (even though it is unknown whether the function of interest satisfies Theorem 1). For example, as shown in Table 4, the theory of the GAW algorithm requires that m should be 4, but this disclosure shows that m=1 can be used. This indicates that in addition to a quadratic speedup in pricing derivatives, further quantum advantage in risk analysis is possible. Classically computing k greeks using finite differences has a complexity of O(k / ε) when the underlying derivatives are priced using Monte Carlo. 2 ), and a straightforward extension of finite difference methods to derivatives pricing using amplitude estimation offers a second order advantage with complexity O(k / ε). In this disclosure, the modified GAW algorithm is

[0419]

number

[0420] Theorem 1 specifies this quadratic advantage in terms of the number of Greeks when a pricing function satisfies the smoothing condition of Theorem 1. However, since derivatives pricing problems of practical interest in finance involve a large number of diverse multivariate price functions and generally do not have analytical solutions, understanding whether and which financial derivatives satisfy the aforementioned smoothing condition can be a challenging task. For this reason, numerical methods were employed to simulate the modified GAW algorithm for two example derivatives, namely, a) European call options, which have a closed-form solution and are used to establish the effectiveness and benchmark of the algorithm, and b) path-dependent basket options, which do not have a known analytical solution and represent a typical derivatives pricing function. The results not only succeed in estimating the relevant Greeks for these examples with high probability, but the resulting query complexity is much smaller than that proposed by Theorem 1 (see, e.g., Tables 3 and 4).

[0421] This disclosure also provides exact resource estimates for the quantum oracle involved in the modified GAW algorithm above. Due to the extra costs associated with the block coding and Hamiltonian simulations used to approximately construct the phase oracle of equation (15) from the probability oracle of equation (1) used in derivatives pricing, Section 5 of this disclosure develops a new method (SFQG method) for constructing a cheaper second-order (m=1) phase oracle by exploiting the structure of the amplitude estimation.

[0422] Section 6 shows that quantum gradient algorithms (e.g., SFQG) can be enhanced by employing maximum likelihood estimation (MLE), making it possible to determine the resources used to estimate the gradient with precise confidence intervals and confidence levels. Using MLE techniques, Section 7 describes estimating the resources used for quantum advantage in derivatives market risk for a typical use case of practical interest. Employing quantum gradient methods in derivatives pricing can reduce the theoretical clock rate estimate for quantum advantage described in Non-Patent Document 3 by a factor of 7.

[0423] Finite difference methods can be used to compute the Greeks, but classical automatic differentiation (AD) techniques can also be used. AD methods can be advantageous due to their ability to significantly reduce the associated computational cost, at the expense of an increased memory footprint. In particular, the adjoint mode of automatic differentiation (AAD) in some cases allows the computation of all k gradients of a scalar function f at a cost that is k-independent, which means that for some constant ω that depends on the function f, the total classical complexity cost in this case is O(ω / ε 2 For example, the complexity of the GAW quantum gradient estimation algorithm is given by

[0424]

number

[0425] Although in section 4.1 we found that for practical use cases in finance, the algorithm scales 100-200 times better than the theoretical estimate from Theorem 1 for a given ε (see, for example, Tables 3 and 4). Thus, depending on the actual scaling of the GAW algorithm to larger values ​​of k for financial use cases, it is also possible that it can outperform the AAD method in complexity. Note that if the GAW algorithm outperforms AD, the SFQG method will outperform as well. However, as shown in the studied examples, the SFQG method is cheaper than GAW, so the SFQG method can outperform AD even in cases where the GAW algorithm does not outperform AD. It is also interesting to consider whether a similar configuration to the one adopted by AD can be applied to the quantum setting. Section 9 provides such a configuration and shows that in some settings it is possible to lead to a performance profile similar to classical AD, in that the running time of the algorithm is independent of the number of Greeks at the expense of increased memory usage.

[0426] Although this disclosure describes calculating slopes, the teachings herein may be applicable to other useful risk metrics, such as the calculation of a portfolio's Value at Risk (VaR). For example, based on the results described herein, further quantum advantages in derivatives pricing subroutines may be leveraged in these higher level calculations and aggregations.

[0427] 11. Exemplary Methods FIG. 7 is a flow chart illustrating a method 700 for determining the gradient of a function f(x) describing an amount of a resource (by SFGQ method), according to one embodiment. For example, the method may be for determining the gradient of a price function of an instrument, such as a financial derivative (e.g., Greeks). The steps of method 700 may be performed in a different order, and method 700 may include different, additional, or fewer steps. In the example of FIG. 7, the steps of method 700 are performed by a computing system (e.g., including a classical computing system (e.g., 810) and a quantum computing system (e.g., 820)).

[0428] The computing system receives 705 a function f(x) describing a quantity of a resource, where x represents a set of k parameters of the quantity. In some aspects, f(x) does not have a closed form solution (e.g., numerical methods are used to approximate f(x)) or f(x) is a function of the Gevrey class G 1 / 2 It is not a function. The function f(x) is computed by using a quantum computing system (e.g., a computing system) to repeatedly execute a quantum function algorithm (e.g., a quantum pricing algorithm) with an error ε p The quantum function algorithm can be determined in O(1 / ε p ) The quantum function algorithm may be a quantum amplitude estimation algorithm.

[0429] The quantum computing system of the computing system repeatedly executes 710 a quantum gradient algorithm to determine a k-dimensional gradient of a function f(x) at a point x0 within error ε. The quantum gradient algorithm uses a phase oracle defined by a finite difference approximation having an order greater than 0.

[0430]

number

[0431] The complexity of the quantum gradient algorithm is

[0432]

number

[0433] In some embodiments, the quantum gradient algorithm scales by:

[0434]

number

[0435] (e.g., m=1 is used instead of m=4 (Theorem 1 requires that m=4 should be used), see Section 5.2, Table 4, and Figure 4). Since smaller values ​​of m can be used, the function is of Gevrey class G 1 / 2 is smoother than the function, and therefore

[0436]

number

[0437] Determining the k-dimensional gradient of the function f(x) may include performing a maximum likelihood estimation (MLE) or performing automatic differentiation.

[0438] Phase Oracle

[0439]

number

[0440] teeth,

[0441]

number

[0442] where m is the finite difference approximation order greater than 0, |x> is a k-dimensional vector representing a set of k parameters, and S is a scaling factor that controls the accuracy of the finite difference approximation.

[0443]

number

[0444] are coefficients of the finite difference approximation of order m. In some embodiments,

[0445]

number

[0446] where c is a Gevrey class G 1 / 2 is the smoothness parameter of the function (this is because the function is of Gevrey class G 1 / 2 (This can occur when m is smoother than the function (defined in Theorem 1). See also Section 5.2, Table 4, and Figure 4, except that m=1 is used instead of m=4 (Theorem 1 requires that m=4 should be used). Smaller values ​​of m can be used so that the function is in the Gevrey class G 1 / 2 is smoother than the function, and therefore

[0447]

number

[0448] For example, m=1.

[0449] In some aspects, executing a quantum gradient algorithm comprises executing a quantum circuit on qubits of a quantum computing system, the executing the quantum circuit comprising: (1) applying a first quantum unitary operator to the qubits in the first register and the second register, the first unitary operator being:

[0450]

number

[0451] where |ψ0(x)〉 and |ψ1(x)〉 are normalized quantum states; and (2) applying a second quantized unitary operator to the qubits in the third register and the fourth register, the second unitary operator being

[0452]

number

[0453] where |ψ0(-x)〉 and |ψ1(-x)〉 are normalized quantum states and may include their application. The first and second unitary operators may be applied in parallel to the respective qubits. Executing the quantum circuit includes (1) applying the first unitary operator followed by applying a third quantum unitary operator to the qubits in the first register and the second register, the third unitary operator being:

[0454]

number

[0455] where:

[0456]

number

[0457] and

[0458]

number

[0459] and (2) applying a fourth quantum unitary operator to the qubits in the third register and the fourth register after applying the second unitary operator, the fourth unitary operator being:

[0460]

number

[0461] and may further include the application of the third and fourth unitary operators. The third and fourth unitary operators may be applied in parallel to the respective qubits. The third unitary operator may be applied multiple times to the qubits in the first and second registers, and the fourth unitary operator may be applied multiple times to the qubits in the third and fourth registers. For example, the third and fourth unitary operators are each applied π*S times, where S=N / Dl. Executing the quantum circuit may further include applying: (1) a Hadamard gate to the qubits in a first register of qubits (e.g., to each qubit in the first register) before applying the first unitary operator; (2) a controlled NOT (CNOT) to the qubits in a third register (e.g., to each qubit in the third register) before applying the second unitary operator, the CNOT gate being controlled by the qubits in the first register; and (3) a CNOT gate to the qubits in the third register (e.g., to each qubit in the third register) after applying the third and fourth unitary operators, the CNOT gate being controlled by the qubits in the first register. Executing the quantum circuit may further include applying an inverse quantum Fourier transform to the qubits in the first register. After the quantum circuit is executed, the quantum state of the qubits on the first register may be measured.

[0462] After the gradient of the function f(x) is determined, the computing system may transmit instructions to buy the resource, sell the resource, or adjust a hedge of a portfolio that includes the resource based on the gradient.

[0463] 12. Description of Computing Systems Aspects described in the above disclosure may be implemented using one or more computing systems. An exemplary computing system is described below.

[0464] FIG. 8A is a block diagram illustrating an embodiment of a computing system 800. In the example of FIG. 8A, the computing system 800 includes a classical computing system 810 (also referred to as a non-quantum computing system) and a quantum computing system 820, although a computing system may include only classical or quantum computing systems. An embodiment of the classical computing system 810 is further described with respect to FIG. 9. Although the classical computing system 810 and the quantum computing system 820 are shown together, they may be physically separate systems. For example, FIG. 8B illustrates an exemplary cloud computing architecture in which the computing system 810 and the quantum computing system 820 communicate via a network 857. The computing system 800 may include different or additional elements (e.g., multiple quantum computing systems 820) than those illustrated. In addition, functionality may be distributed among the elements in a manner different than that described.

[0465] The classical computing system 810 may control the quantum computing system 820. For example, the classical computing system 810 generates and sends instructions for the quantum computing system 820 to execute a quantum algorithm or quantum circuit. Although only one classical computing system 810 is shown in FIG. 8A , any number of classical computing systems 810 or other external systems may be connected to the quantum computing system 820.

[0466] FIG. 8C is a block diagram illustrating one embodiment of a quantum computing system 820. The quantum computing system 820 includes a number of quantum bits ("qubits") 850 and an associated quantum bit controller 840. As shown in FIG. 8D, the quantum bits 850 may be in a quantum bit register (or multiple registers) of the quantum computing system 820. The quantum bits are further described below. The quantum bit controller 840 is a module that controls one or more quantum bits 850. The quantum bit controller 840 may include a classical processor such as a CPU, GPU, or FPGA. The quantum bit controller 840 may perform physical operations on one or more quantum bits 850 (e.g., it may perform quantum gate operations on the quantum bits 840). Although a separate quantum bit controller 840 is shown for each quantum bit 850 in the example of FIG. 8C, the quantum bit controller 850 may control multiple (e.g., all) quantum bits 850 of the quantum computing system 820, or multiple controllers 850 may control a single quantum bit. For example, qubit controller 850 may be a separate processor, parallel threads on the same processor, or some combination of both.

[0467] FIG. 8E is a flow chart illustrating an exemplary execution of a quantum routine on a computing system 800. The classical computing system 810 generates a quantum program (860) that is executed or processed by the quantum computing system 820. The quantum program may include instructions or subroutines that are executed by the quantum computing system 820. In one example, the quantum program is a quantum circuit. The quantum computing system 820 executes the program (865) and calculates results (called shots or runs) (870). Computing the results may include performing measurements of the quantum states generated by the quantum computing system 820 that resulted from executing the program. In practice, this may be performed by measuring the values ​​of one or more of the qubits 850. The quantum computing system 820 typically performs multiple shots to accumulate statistics from the stochastic execution. The number of shots and any changes (e.g., parameter changes) that occur between shots may be referred to as a schedule. The schedule may be specified by the program. The results (e.g., quantum state data) (or accumulated results) are recorded (875) by the classical computing system 810. The results may be returned after a termination condition is met (e.g., a threshold number of shots has occurred). The classical computing system 810 may determine the quantity based on the received results.

[0468] Quantum computing system 820 exploits the laws of quantum mechanics to perform calculations. A quantum processing device (QPU), a quantum computer, a quantum processor, and a quantum processing unit are each examples of a quantum computing system. Quantum computing system 800 can be a universal or non-universal quantum computing system (a universal quantum computing system can execute any possible quantum circuit (subject to the constraint that the circuit does not use more qubits than the quantum computing system)). In some aspects, quantum computing system 800 is a gate model quantum computer. As mentioned above, quantum computing systems use so-called qubits or quantum bits (e.g., 850A). While classical bits always have a value of either 0 or 1, quantum bits are quantum mechanical systems that can have a value of 0, 1, or a superposition of both values. Exemplary physical implementations of quantum bits include superconducting qubits, spin qubits, trapped ions, arrays of neutral atoms, and photonic systems (e.g., photons in a waveguide). For purposes of this disclosure, a qubit may be realized by a single physical qubit or as an error-protected logical qubit that itself includes multiple physical qubits. In addition, this disclosure is not specific to qubits. This disclosure may be generalized to apply to quantum computing systems 820 whose building blocks are qubits (d-level quantum systems, where d>2) or quantum continuous variables rather than qubits.

[0469] A quantum circuit is an ordered set of one or more gates. A subcircuit may refer to a circuit that is part of a larger circuit. A gate represents a unitary operation performed on one or more quantum bits. A quantum gate may be described using a unitary matrix. The depth of a quantum circuit is the minimum number of steps used to execute the circuit on a quantum computing system. The depth of a quantum circuit may be less than the total number of gates, since gates that operate on non-overlapping subsets of quantum bits may be executed in parallel. A layer of a quantum circuit may refer to a step of a circuit during which multiple gates may be executed in parallel. In some aspects, a quantum circuit is executed by a quantum computing system. In this sense, a quantum circuit may be thought of as including a set of instructions or operations that a quantum computing system can execute. To execute a quantum circuit on a quantum computing system, a user may inform the quantum computing system which circuits should be executed. A quantum computing system may include both a core quantum device and classical peripheral / control devices (e.g., qubit controller 840) used to coordinate the control of the quantum device. It is to this classical control device that a description of the quantum circuit can be sent when trying to get a quantum computer to execute the circuit.

[0470] The parameters of a parameterized quantum circuit may refer to the parameters of the gates. For example, a gate that performs a rotation around the y-axis may be parameterized by a real number that describes the angle of the rotation.

[0471] Descriptions of quantum circuits executed on one or more quantum computing systems may be stored on a non-transitory computer-readable recording medium. The term "computer-readable recording medium" should be considered to include a single medium or multiple media capable of storing instructions (e.g., a centralized or distributed database, or associated caches and servers). The term "computer-readable medium" should also be considered to include any medium capable of storing instructions for execution by a quantum computing system, causing the quantum computing system to perform any one or more of the methods disclosed herein. The term "computer-readable medium" includes, but is not limited to, data repositories in the form of solid-state memory, optical media, and magnetic media.

[0472] FIG. 9 is an example architecture of a classical computing system 810, according to one embodiment. A quantum computing system 820 may also have one or more of the components described with respect to FIG. 9. FIG. 9 illustrates a high-level block diagram showing the physical components of a computer system used as part or all of one or more entities described herein, according to one embodiment. A computer may have additional, fewer, or variations of the components provided in FIG. 9. Although FIG. 9 illustrates a computer 900, this diagram is intended as a functional description of various features that may be present in a computer system, rather than as a structural schematic of the implementations described herein. In practice, items shown separately may be combined and some items may be separated, as will be recognized by those skilled in the art.

[0473] 9 illustrates at least one processor 902 coupled to a chipset 904. Also coupled to the chipset 904 are memory 906, storage device 908, keyboard 910, graphics adapter 912, pointing device 914, and network adapter 916. A display 918 is coupled to the graphics adapter 912. In one aspect, the functionality of the chipset 904 is provided by a memory controller hub 920 and an I / O hub 922. In another aspect, the memory 906 is directly coupled to the processor 902 instead of the chipset 904. In some aspects, the computer 900 includes one or more communication buses for interconnecting these components. The one or more communication buses optionally include circuitry (sometimes referred to as a chipset) that interconnects and controls communication between the system components.

[0474] The storage device 908 is any non-transitory computer-readable recording medium, such as a hard drive, a compact disk read-only memory (CD-ROM), a DVD, or a solid-state memory device or other optical storage device, a magnetic cassette, a magnetic tape, a magnetic disk storage device or other magnetic storage device, a magnetic disk storage device, an optical disk storage device, a flash memory device, or other non-volatile solid-state storage device. Such storage device 908 may also be referred to as persistent memory. The pointing device 914 may be a mouse, a trackball, or other type of pointing device, and is used in combination with the keyboard 910 to input data into the computer 900. The graphics adapter 912 displays images and other information on the display 918. The network adapter 916 couples the computer 900 to a local or wide area network.

[0475] The memory 906 holds instructions and data used by the processor 902. The memory 906 can be a non-persistent memory, examples of which include high-speed random access memory such as DRAM, SRAM, DDR RAM, ROM, EEPROM, flash memory, etc.

[0476] As is known in the art, computer 900 may have different components or other components than those shown in FIG. 9. Additionally, computer 900 may lack some illustrated components. In one aspect, a computer 900 functioning as a server may lack a keyboard 910, a pointing device 914, a graphics adapter 912, or a display 918. Additionally, storage device 908 may be local or remote from computer 900 (such as embodied in a storage area network (SAN)).

[0477] As known in the art, the computer 900 is adapted to execute computer program modules for providing the functionality described herein. The term "module" as used herein refers to computer program logic utilized to provide a specified functionality. Thus, a module can be implemented in hardware, firmware, or software. In one aspect, the program module is stored in the storage device 908, loaded into the memory 906, and executed by the processor 302.

[0478] 13. Additional Considerations The above disclosure describes exemplary embodiments for purposes of illustration only, and any features described as essential, critical, or otherwise implied to be required should be construed as required for that embodiment only, and not necessarily included in other embodiments.

[0479] In addition, the above disclosure uses "we" (and other similar terms) to refer to entities performing operations (e.g., steps in an algorithm). These terms are used for convenience. These terms may refer to a computing system (e.g., computing system 100) that is performing the operations described.

[0480] Some portions of the above disclosure describe aspects related to algorithmic processes or operations. These algorithmic descriptions and representations are commonly used by those skilled in the computing arts to effectively convey the substance of their work to others skilled in the art. These operations, while described functionally, computationally, or logically, will be understood to be implemented by computer programs including instructions for execution by a processor or equivalent electrical circuitry, microcode, or the like. Moreover, it has also proven convenient at times to refer to these configurations of functional operations as modules, without loss of generality. In some cases, modules may be implemented in hardware, firmware, or software.

[0481] Any reference to "one embodiment" or "an embodiment" as used herein means that a particular element, feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. The appearances of the phrase "in one embodiment" in various places in this specification do not necessarily all refer to the same embodiment. Similarly, the use of "a" or "an" preceding an element or component is done for convenience only. The description should be understood to mean that there are one or more elements or components, unless it is clear that something else is meant. As used herein, the terms "comprises," "comprising," "includes," "including," "has," "having," or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, article, or apparatus comprising a list of elements is not necessarily limited to only those elements, and may include other elements not expressly listed or other elements not inherent to such process, method, article, or apparatus. Furthermore, unless expressly stated to the contrary, "or" refers to an inclusive "or" and not an exclusive "or." For example, a condition A or B is satisfied by any one of the following: A is true (or exists) and B is false (or does not exist), A is false (or does not exist) and B is true (or exists), and both A and B are true (or exist).

[0482] In addition, the use of "a" or "an" is employed to describe elements and components of an embodiment. This is done merely for convenience and to give a general sense of the disclosure. This description should be read to include one or at least one, and the singular also includes the plural, unless it is clear that it is meant otherwise. When a value is described as "about" or "substantially" (or derivatives thereof), such value should be interpreted as being accurate to + / - 10%, unless another meaning is clear from the context. For example, "about 10" should be understood to mean "in the range of 9 to 11".

[0483] Alternative aspects are implemented in computer hardware, firmware, software, and / or combinations thereof. The implementation may be implemented in a computer program product tangibly embodied in a machine-readable storage device for execution by a programmable processor, and the method steps may be performed by the programmable processor executing a program of instructions to perform functions by operating on input data and generating output. As used herein, a "processor" may refer to one or more processors. The aspects may be advantageously implemented in one or more computer programs executable on a programmable system including at least one programmable processor coupled to receive data and instructions from, and transmit data and instructions to, a data storage system, at least one input device, and at least one output device. Each computer program may be implemented in a high-level procedural or object-oriented programming language, or in assembly or machine language as appropriate, and in either case the language may be a compiled or interpreted language. Suitable processors include, by way of example, both general-purpose and special-purpose microprocessors. In general, the processor receives instructions and data from a read-only memory and / or a random access memory. Typically, a computer will include one or more mass storage devices for storing data files; such devices include magnetic disks, such as internal hard disks and removable disks, magneto-optical disks, and optical disks. Storage devices suitable for tangibly embodying computer program instructions and data include all forms of non-volatile memory, including, by way of example, semiconductor memory devices, such as EPROM, EEPROM, and flash memory devices, magnetic disks, such as internal hard disks and removable disks, magneto-optical disks, and CD-ROM disks. Any of the foregoing may be supplemented by, or incorporated in, ASICs (application-specific integrated circuits) and other forms of hardware.

[0484] Although the above description contains many details, these should not be interpreted as limiting the scope of the present disclosure, but merely as illustrating different examples. It should be appreciated that the scope of the present disclosure includes other aspects not described in detail above. Various other modifications, changes, and variations that will be apparent to those skilled in the art may be made in the configuration, operation, and details of the methods and apparatus disclosed herein without departing from the spirit and scope of the present disclosure.

Claims

1. Receiving a function f(x) describing a quantity of a resource, where x represents a set of k parameters of said quantity; 2. Iteratively execute the quantum gradient algorithm of the quantum computing system to find the point x 0 determining a k-dimensional gradient for the function f(x) within an error ε in sf m The complexity of the quantum gradient algorithm is [0010] Scaling as 23. A method comprising:

2. The phase oracle O sf m teeth, [0025] where m is the order of the finite difference approximation greater than 0, x|> is a k-dimensional vector representing the set of k parameters, S is a scaling factor that controls the accuracy of the finite difference approximation, and a l (2m) are the coefficients of the finite difference approximation of order m 2. The method of claim 1 .

3. [Math. 3] where c is the Gevrey class G 1/2 3. The method of claim 2, wherein the parameters of the function are smooth.

4. 4. The method of claim 3, wherein m=3.

5. Repeatedly executing the quantum function algorithm of the quantum computing system to obtain an error ε p 2. The method of claim 1, further comprising determining the function f(x) within a range:

6. 6. The method of claim 5, wherein the quantum function algorithm is a quantum amplitude estimation algorithm.

7. The complexity of the quantum function algorithm is O(1 / ε p 6. The method of claim 5, wherein the scaling is performed as

8. 2. The method of claim 1, wherein f(x) does not have a closed form solution.

9. f(x) is a Gevrey class G 1/2 2. The method of claim 1, wherein the function is not a function.

10. 2. The method of claim 1, wherein executing the quantum gradient algorithm comprises executing a quantum circuit on qubits for the quantum computing system.

11. Executing the quantum circuit includes: applying a first quantum unitary operator to qubits in a first register and in a second register, the first unitary operator being: [0045] where |Ψ 0 (x)> and |Ψ 1 (x)> is a normalized quantum state; and applying a second quantum unitary operator to the qubits in the third register and in the fourth register, the second unitary operator being [0050] where |Ψ 0 (−x)> and |Ψ 1 (-x)> is a normalized quantum state, 11. The method of claim 10, comprising:

12. Executing the quantum circuit includes: subsequent to applying the first unitary operator, applying a third quantum unitary operator to qubits in the first register and in the second register, the third unitary operator being: [006] where: [0070] and [0080] That is, subsequent to applying the second unitary operator, applying a fourth quantum unitary operator to qubits in the third register and in the fourth register, the fourth unitary operator being: [0090] and The method of claim 11 further comprising:

13. the third unitary operator is applied multiple times to qubits in the first register and in the second register; The fourth unitary operator is applied multiple times to qubits in the third register and in the fourth register.

13. The method of claim 12.

14. Executing the quantum circuit includes: applying a Hadamard gate to qubits in the first register of qubits prior to applying the first unitary operator; applying a controlled NOT (CNOT) gate to a qubit in the third register prior to applying the second unitary operator, the CNOT gate being controlled by a qubit in the first register; subsequent to applying the third and fourth unitary operators, applying a CNOT gate to a qubit in the third register, the CNOT gate being controlled by a qubit in the first register; The method of claim 13 further comprising:

15. 15. The method of claim 14, wherein executing the quantum circuit further comprises applying an inverse quantum Fourier transform to qubits in the first register.

16. The quantum gradient algorithm is [0010] 2. The method of claim 1, wherein the method scales better than

17. 2. The method of claim 1, wherein determining the k-dimensional gradient for the function f(x) comprises performing a maximum likelihood estimation (MLE).

18. 2. The method of claim 1, wherein determining the k-dimensional gradient for the function f(x) comprises performing automatic differentiation.

19. Receiving a function f(x) describing a quantity of a resource, where x represents a set of k parameters of said quantity; providing a quantum circuit configured to implement a quantum gradient algorithm when executed on a four qubit register of a quantum computing system, the quantum gradient algorithm comprising: ##EQU00011## and the phase oracle O defined by a finite difference approximation with order greater than 0. sf m The quantum circuit includes the phase oracle O sf m(1) a first quantum unitary operator to be applied to qubits in the first and second registers; and (2) a second quantum unitary operator to be applied to qubits in the third and fourth registers to implement Executing the quantum circuit repeatedly on the quantum bits of the quantum computing system to obtain a point x 0 determining a k-dimensional gradient for the function f(x) within an error ε in 23. A method comprising:

20. 20. The method of claim 19 further comprising the limitations of any of claims 2 to 18.

21. receiving a function f(x) describing a quantity of a resource, where x is a k-dimensional vector representing a set of k parameters of said quantity; generating a set of instructions for repeatedly executing a quantum gradient algorithm for a quantum computing system, the quantum gradient algorithm comprising a phase oracle O defined by a finite difference approximation having an order greater than 0; sf m The complexity of the quantum gradient algorithm is ##EQU00012## Scaling as transmitting the instruction to the quantum computing system; receiving quantum state data from the quantum computing system; Based on the quantum state data, point x 0 determining a k-dimensional gradient for the function f(x) within an error ε in 23. A method comprising:

22. 22. The method of claim 21 further comprising any of the limitations of claims 2 to 18.

23. When executed by a computing system, the computing system is Receiving a function f(x) describing a quantity of a resource, where x represents a set of k parameters of said quantity; 2. Iteratively execute the quantum gradient algorithm of the quantum computing system to find the point x 0 determining a k-dimensional gradient for the function f(x) within an error ε in sf m The complexity of the quantum gradient algorithm is ##EQU00013## Scaling as A non-transitory computer-readable storage medium comprising stored instructions that cause a computer to perform operations including:

24. 24. The non-transitory computer-readable storage medium of claim 23, further comprising stored instructions that, when executed by a computing system, cause the computing system to perform the operations of any of the limitations of claims 2 to 18.

25. a quantum computing system configured to execute a quantum circuit; receiving a function f(x) describing a quantity of a resource, where x is a k-dimensional vector representing a set of k parameters of said quantity; generating a set of instructions for repeatedly executing a quantum gradient algorithm for a quantum computing system, the quantum gradient algorithm comprising a phase oracle O defined by a finite difference approximation having an order greater than 0; sf m The complexity of the quantum gradient algorithm is ##EQU14## Scaling as transmitting the instruction to the quantum computing system; receiving quantum state data from the quantum computing system; Based on the quantum state data, point x 0 determining a k-dimensional gradient for the function f(x) within an error ε in a classical computing system configured to perform operations including: A system comprising:

Citation Information

Patent Citations

  • Estimating the energy levels of physical systems

    JP2021524622A

  • Hybrid quantum-classical computer for bayesian inference with engineered likelihood functions for robust amplitude estimation

    WO2020252425A1