Quantum Metropolis-Hastings algorithm with the target distribution calculated by quantum Monte Carlo integration

Koichi Miyamoto

2023-03-10

Abstract:The Markov chain Monte Carlo method (MCMC), especially the Metropolis-Hastings (MH) algorithm, is a widely used technique for sampling from a target probability distribution $P$ on a state space $\Omega$ and applied to various problems such as estimation of parameters in statistical models in the Bayesian approach. Quantum algorithms for MCMC have been proposed, yielding the quadratic speedup with respect to the spectral gap $\Delta$ compered to classical counterparts. In this paper, we consider the quantum version of the MH algorithm in the case that calculating $P$ is costly because the log-likelihood $L$ for a state $x\in\Omega$ is obtained via computing the sum of many terms $\frac{1}{M}\sum_{i=0}^{M-1} \ell(i,x)$. We propose calculating $L$ by quantum Monte Carlo integration and combine it with the existing method called quantum simulated annealing (QSA) to generate the quantum state that encodes $P$ in amplitudes. We consider not only state generation but also finding a credible interval for a parameter, a common task in Bayesian inference. In the proposed method for credible interval calculation, the number of queries to the quantum circuit to compute $\ell$ scales on $\Delta$, the required accuracy $\epsilon$ and the standard deviation $\sigma$ of $\ell$ as $\tilde{O}(\sigma/\epsilon^2\Delta^{3/2})$, in contrast to $\tilde{O}(M/\epsilon\Delta^{1/2})$ for QSA with $L$ calculated exactly. Therefore, the proposed method is advantageous if $\sigma$ scales on $M$ sublinearly. As one such example, we consider parameter estimation in a gravitational wave experiment, where $\sigma=O(M^{1/2})$.

Quantum Physics,Instrumentation and Methods for Astrophysics,General Relativity and Quantum Cosmology

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to accelerate the calculation of the target distribution \(P\) in the quantum computing environment when the calculation cost of the target distribution \(P\) in the Metropolis - Hastings (MH) algorithm is high, especially when the negative log - likelihood \(L(x)\) is the sum of many terms. Specifically, the author focuses on the problem that in Bayesian inference, when \(P(x)\propto P_0(x)e^{-L(x)}\) and \(L(x)=\frac{1}{M}\sum_{i = 0}^{M - 1}\ell(i,x)\), calculating \(L(x)\) requires a large number of iterations, resulting in an overly long calculation time.

To solve this problem, the author proposes a method that combines quantum Monte Carlo integration (QMCI) and quantum simulated annealing (QSA). The specific steps are as follows:

1. **Estimate \(L(x)\) using quantum Monte Carlo integration (QMCI)**: Estimate \(L(x)\) through QMCI, thereby reducing the time required to calculate \(L(x)\). QMCI provides a quadratic acceleration relative to classical Monte Carlo integration.

2. **Generate the quantum state \(|P\rangle\) by combining quantum simulated annealing (QSA)**: Use the QSA method to gradually generate the quantum state \(|P\rangle\) that encodes the target distribution \(P\). The QSA method approximates the final target distribution \(P\) through a series of intermediate distributions \(P_{\beta_i}\).

3. **Calculate the confidence interval**: Not only generate the quantum state \(|P\rangle\), but also propose a method for calculating the confidence interval of parameters. Iteratively use QMCI to estimate the cumulative distribution function (CDF), and then find the confidence interval through binary search.

The advantage of this method is that when \(\sigma\) (i.e., the standard deviation of \(\ell\)) grows sub - linearly with respect to \(M\), the complexity is better than that of the classical MH algorithm and the exact QSA method. Especially in the example of gravitational - wave parameter estimation, \(\sigma = O(\sqrt{M})\), so this method has a quadratic acceleration effect in terms of \(M\).

In summary, this paper aims to provide a quantum algorithm for efficiently calculating the target distribution \(P\) by combining QMCI and QSA, which is especially suitable for scenarios that require a large amount of calculation of \(L(x)\), such as gravitational - wave parameter estimation in Bayesian inference.