The present invention relates to a computing device for an Ising model, and more particularly, to a probabilistic distribution of spin states of an Ising model generated in a computing device.
The Ising model is a numerical model in which a large number (N) of spins represented by a discrete variable σ having a binary value of (−1, 1) are coupled on a network, and a specific spin state σ=(σ1, . . . , σN)T including N spins and an energy E(σ) of the state can be associated as in the following Formula (1).
Here, J is a spin adjacency matrix of N rows and N columns, and each element of the matrix represents a coupling constant between spins on the network or an interaction between spins. The Ising model was originally considered in physics as a model representing thermodynamic characteristics of a ferromagnet. Then, with the development of electronic computers, the application range has been expanded, such as being used to express a model representing a discrete optimization problem or a learning model in statistical machine learning due to the simplicity and ability to increase the scope of the model.
Here, as thermodynamic characteristics, it is known that the properties of the Ising model in a thermal equilibrium state can be approximated by assuming that a microscopic spin state stochastically exists in a canonical distribution as expressed by the following Formula (2).
Here, β is the reciprocal (inverse temperature) of an absolute temperature when the Boltzmann constant in the thermal equilibrium system is 1, and Z is
On the other hand, in a non-equilibrium state out of the thermal equilibrium state, a distribution that may typically occur in different systems is known. This is, for example, a power law distribution which is characterized by a q-exponential function using an Ising energy as an argument in Tsallis statistics (hereinafter, also referred to as Tsallis distribution) (refer to Non Patent Literature 1). In the case of the Ising model having the Tsallis distribution, the probability distribution of the spin state is given by the following Formula (3).
Here, expq(x)≡[1+(1−q)x]1/(1-q) is a q-exponential function obtained by extending a normal exponential function with a real number parameter q, and Zq is
As described above, a power law distribution such as a Tsallis distribution has a property of gradual decay with respect to energy, which is useful for sampling a spin state. Therefore, if an Ising model having such a power law distribution can be realized by a computing device, there is a possibility that it will be useful for sampling of a spin state in an application range such as machine learning.
An object of the present invention is to provide a computing device that generates a spin state of an Ising model according to a distribution that can be approximated by a power law distribution and samples the spin state.
In order to achieve such an object, a computing device of one aspect of the present invention includes a means for generating a spin state for which a distribution indicating a probability of existence of the spin state for each energy of the spin state is available to be approximated by a power law distribution.
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
Embodiments of the present invention relate to a coherent Ising machine. Specifically, as a computing device that generates a spin state of an Ising model having a Tsallis distribution, an embodiment (embodiment 1) in which a Tsallis distribution is generated by adjusting pump light in a non-oscillation region in the vicinity of an oscillation point in an actual coherent Ising machine and an embodiment (embodiment 2) in which the Tsallis distribution is generated by simulating an operation of the actual machine on an electronic computer will be described.
In
As described above, an interaction (coupling constant) between pulses (that is, between spins) is given to the light pulses circulating in the ring resonator 1 by performing measurement and feedback. That is, an optical coupler 5 extracts a part of each light pulse for each circulation of the light pulses. Then, a measurement unit 6 measures a signed amplitude xj of a light pulse in the phase 0 or π direction. The sign is positive in the phase 0 direction and negative in the phase π direction. An arithmetic unit 7 is given in advance information of a coupling constant between adjacency matrices, and calculates a coupling signal to be fed back and input by using a measurement result of the measurement unit 6. A coupling signal Δxj fed back to the i-th light pulse is expressed by the following Formula (4).
Here, ξ is a dimensionless feedback intensity. Next, an external light pulse input unit 8 generates an external light pulse according to the calculated coupling signal Δxj, and inputs the external light pulse into the ring resonator 1 via the optical coupler 9. By measurement/feedback control described above, it is possible to give a mutual relationship to phases between light pulses constituting a light pulse train. The light pulse train is circularly amplified in the ring resonator 1 while this correlation is given, and the phases 0 and π of the light pulses constituting the light pulse train are measured when a stable state, that is, a steady state in which a state change of the light pulses is substantially eliminated is obtained by advancing time evolution. This final measurement result is sent to an electronic computer (not illustrated) and used for a desired purpose.
In the conventional coherent Ising machine, measurement is performed with the amplitude of the pump light 3 larger than an oscillation point, that is, the oscillation threshold value (amplitude) at the time of measuring a phase in a stable state (Non Patent Literature 3).
On the other hand, in the embodiment of the present invention, the amplitude of the pump light 3 is maintained at the oscillation point or less, and measurement is performed in a state where the value is maintained. As a result, the influence of the quantum fluctuation can be increased, and the spin state of the Ising model according to the power law distribution that can be approximated by the Tsallis distribution can be sampled. An exponent γ=1/(q−1) representing a decay of probability at high energy varies with pump light intensity. As the pump light intensity approaches the oscillation point from zero, it monotonically increases from γ=0 to γ=γ0>0. The value of the maximum value γ0 of the exponent depends on a given coupling coefficient matrix. Therefore, it is possible to generate a power law distribution in which the exponent has been changed relatively freely by changing the pump light intensity. Here, the distribution is a power law distribution indicating the probability that the spin state (+1 or −1, 0 or π in the present embodiment) exists for each energy of the spin state, and in the present embodiment, this distribution is a power law distribution that can be approximated by the Tsallis distribution.
Specifically, first, in the present embodiment, the phase of a light pulse is set to 0 or π and the phase 0 or π is measured in a state where the amplitude of the pump light 3 is set to a value equal to or less than the oscillation point. This is because even if the amplitude of the pump light is equal to or less than the oscillation point, when the pump light amplitude is close to the vicinity of the oscillation point, fluctuation grows in a specific direction of 0 or π, and the amplitude of the pulse is subject to strong quantum fluctuation distorted in this specific direction, and the probability that the phase takes a value in the direction of 0 or π increases. After oscillation, the quantum mechanical fluctuation effect is weakened, and the phase of 0 or π is determined. Further, also in measurement of the phase, as described above, the amplitude projected in the direction of 0 or π is measured and only information in the direction of 0 or π is acquired. As a result, phase information obtained by the measurement becomes a reduced binary value of 0 or n.
Next, in the present embodiment, a phase state (spin state) of light pulses in the power law distribution represented by the Tsallis distribution is generated in a state where the amplitude of the pump light is set to a value equal to or less than the oscillation point. This distribution generation was found by the inventors of the present application in numerical experiments using numerical computation. As a result, it is possible to sample the phase state of each light pulse in the phase state of the power law distribution.
As described above, the present embodiment realizes a computing device of an Ising model with a power law distribution (Tsallis distribution). As a result, the property of gradual power low decay with respect to energy may be useful in sampling of a spin state in an application range such as machine learning. This is because the power law distribution approximates an actual existing distribution such as in a physical phenomenon, an economic phenomenon, a social phenomenon. The canonical distribution used in the conventional Ising model computing devices is a kind of distribution called an exponential distribution family, and distinguishes the appearance probability of each spin state by an exponential function of energy. That is, it is a distribution in which a low energy state likely to occur is distinguishable from a high energy state less likely to occur and negligible. On the other hand, in the power law distribution, the appearance probability of a spin state decreases relatively slowly rather than exponentially as the energy state increases. Therefore, the appearance probability cannot be distinguished so clearly. It is known that such a power law distribution is a distribution that is often seen in an actual physical phenomenon or the like. The present embodiment can be used for applications such as machine learning by generating this power law distribution in the Ising model.
The configuration illustrated in
Therefore, by using the branching ratio variable optical coupler 50, the branching ratio is maintained at a low value during a normal operation, the branching ratio is changed when the distribution is measured, and, for example, all photons (100%) are sent to the measurement unit. Note that the response time of the branching ratio variable optical coupler 50 needs to be sufficiently shorter than the time required for one round of each light pulse. By using such a branching ratio variable optical coupler, it is possible to more accurately measure a phase state (spin state) of a light pulse.
The configuration illustrated in
In the embodiment described above, it is necessary to accurately know an oscillation point in order to fix the pump light amplitude to the oscillation point or a value slightly less than the oscillation point. On the other hand, this value of the oscillation point changes by changing an adjacency matrix of the Ising model. Therefore, it is possible to experimentally search for the oscillation point by gradually changing the pump light intensity every time the adjacency matrix is changed, but it is also possible to input a predetermined value in advance to the diagonal term (self-feedback term, which is usually zero) of the adjacency matrix such that the oscillation point does not change. Specifically, the value of the oscillation point can be fixed by correcting the values of all diagonal terms in advance as expressed by the following Formula (5).
Here, λ0 is the maximum eigenvalue of the original adjacency matrix J.
In another embodiment of the present invention, a distribution of phases (spin states) of the light pulses can also be computed by simulating the coherent Ising machine described in embodiment 1 and the modified example thereof described above on an electronic computer. A specific stochastic differential equation implemented on the electronic computer is given by the following Formula (6).
Here, Q≡(X, P)T is a complex vector variable having 2N elements (X and P each have N elements), and physically corresponds to an amplitude operator represented by normal order of two types of orthogonal phases of each light pulse. τ is a time normalized with a loss rate of a ring resonator, and Wτ is a real number vector having 2N independent Wiener processes as elements when viewed at the time τ. The vector value function f(Q) is given by the following Formula (7).
Further, G is a matrix and satisfies the following Formula (8).
Here, IN appearing in Formulas (7) and (8) is an identity matrix with N rows and N columns, and DIAG (A) is a diagonal matrix whose diagonal elements match diagonal elements of a matrix A. p is a pump light amplitude normalized such that the oscillation point when there is no coupling is 1, and ξ is the feedback intensity expressed by Formula (4). μ is a loss rate by measurement normalized by the loss rate of the ring resonator, and takes a value in a range of 0 (sending 0% of all photons to the measurement unit) to 1 (sending 100% of all photons to the measurement unit). Further, σg corresponds to the reciprocal of the square root of a photon number expectation value in each light pulse in a case where there is no feedback coupling and the pump light amplitude is set to five times the oscillation point (ξ=0, p=5).
The time evolution of the above equation is computed on an electronic computer using the Euler-Maruyama method. At this time, G(Q) is computed for each time evolution step from Formula (8) using numerical Takagi decomposition of the matrix (refer to Non Patent Literature 4). In this manner, the complex variable vector Q is time-evolved from the time τ=0 to a measurement time τ=T. Furthermore, this time evolution is repeated a plurality of times (1000 to 20,000 times), and a spin state distribution P(σ) that may be measured by the measurement unit (6, 60, 600) can be computed as represented by Formula (9).
Here, erf(z) is an error function, and σm≡σg/r is a real number value determined by the aforementioned σg and a branching ratio r (0≤r≤1) in a direction of an optical coupler to the measurement unit at the time of distribution measurement. Further, EQ[h(Q)] represents an operation of averaging the function h(Q) using a plurality of Q values obtained by time evolution computation.
Note that, as the stochastic differential equation, an approximate equation in which a specific term of Formulas (6), (7), and (8) is ignored, an approximate equation modified such that any element of the real part or the imaginary part of the complex vector variable Q is ignored, or an approximate equation obtained by combining them may be used.
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/JP2022/008378 | 2/28/2022 | WO |