https://arxiv.org/abs/1908.03740 (or view on SciRate)

- A QMC algorithm is given by using the “permutation matrix representation” (PMR), which writes the Hamiltonian in two parts, a diagonal part and off-diagonal \(H=D_0 + \sum_i D_i P_i\) where \(D_i\) are diagonal, \(P_i\) off-diagonal permutation matrices \(P_i \vert z\rangle =\vert z'\rangle\) (\(z\neq z'\)).
- The partition function is expanded as a Taylor series, but rewritten as a sum over \(q\), the number of permutations \(\prod_{i=1}^q P_i\)
- This gives a QMC algorithm which in principle works in all parameter ranges (from quantum to classical)

- The goal of Quantum Monte Carlo (QMC) is to obtain samples according to the distribution \(\exp (-\beta H)\), where \(H\) is a quantum Hamiltonian (i.e. one with non-zero off-diagonal components in your basis of choice)
- QMC algorithms write the partition function as a sum of positive weights: \(Z = \mathrm{Tr} e^{-\beta H} = \sum_c W_c\)
- Sampling this system only requires one to propose configuration updates \(c\rightarrow c'\), and if the weights are indeed all positive \(W_c > 0\), the transition probability is computed from the ratio \(W_{c'}/W_{c}\), and standard MC methods can be employed (if any weight is negative, the so-called sign problem exists and standard methods can not be used)
- There are many QMC algorithms around, and most have limits under which they operate well (such as only at high temperature, or at high or low transverse-field)
- The technique used here is perhaps most closely related to the SSE (stochastic series expansion) formulation of QMC, which also starts with a Taylor series for \(Z\) and regroups terms so it is no longer a high temperature expansion
- This work builds directly on arXiv:1701.01499 and arXiv:1802.08333 which proposed splitting the Hamiltonian in a similar manner, but only using generic off-diagonal terms \(H = D_0 + \sum_i V_i\)

- The main contribution of this work over the previous is that it explicitly expands \(H\) using permutation matrices, which offers new insight
- This gives a clean interpretation of the QMC configuration, as a sequence of permutations \(\vert z_0\rangle \rightarrow \vert z_1\rangle \rightarrow \dots\), akin to what is present in the path-integral formulation (using imaginary time-slices)
- The update scheme can dynamically change the \(q\) value (this parameter is similar to the number of time slices in the path-integral approach, which is usually fixed in that case)
- There is a central object called the exponent of divided differences present in the QMC weights, which has some nice properties
- Updates here can in general be costly \(O(q^2)\) as one needs to calculate the exponent of divided differences, however local updates can be done in \(O(q)\)
- Several comparisons against the SSE are given, showing the PMR outperforming in all cases
- The emergence is the sign problem is discussed (nb. there is a follow up paper with more detail in arXiv:2012.02022)

tags:

stoquastic

QMC