Perils of Embedding for Quantum Sampling

TL;DR

This paper investigates the limitations of minor embedding in quantum sampling, especially in the transverse-field Ising model, revealing biases and shifts in phase transitions through advanced quantum Monte Carlo simulations.

Contribution

It generalizes previous work by analyzing quantum thermal sampling with non-zero off-diagonal terms and introduces an improved QMC method for larger system simulations.

Findings

Embedding biases affect observable measurements.

Critical point shifts with embedding size.

Sampling probability depends on system parameters.

Abstract

Given quantum hardware that enables sampling from a family of natively implemented Hamiltonians, how well can one use that hardware to sample from a Hamiltonian outside that family? A common approach is to minor embed the desired Hamiltonian in a native Hamiltonian. In Phys. Rev. Research 2, 023020 (2020) it was shown that minor embedding can be detrimental for classical thermal sampling. Here, we generalize these results by considering quantum thermal sampling in the transverse-field Ising model, i.e. sampling a Hamiltonian with non-zero off diagonal terms. To study these systems numerically we introduce a modification to standard cluster update quantum Monte-Carlo (QMC) techniques, which allows us to much more efficiently obtain thermal samples of an embedded Hamiltonian, enabling us to simulate systems of much larger sizes and larger transverse-field strengths than would otherwise be possible. Our numerics focus on models that can be implemented on current quantum devices using planar two-dimensional lattices, which exhibit finite-temperature quantum phase transitions. Our results include: i) An estimate on the probability to sample the logical subspace directly as a function of transverse-field, temperature, and total system size, which agrees with QMC simulations. ii) We show that typically measured observables (diagonal energy and magnetization) are biased by the embedding process, in the regime of intermediate transverse field strength, meaning that the extracted values are not the same as in the native model. iii) By considering individual embedding realizations akin to 'realizations of disorder', we provide numerical evidence suggesting that as the embedding size is increased, the critical point shifts to increasingly large values of the transverse-field.