SciPost Submission Page
Importance Sampling Scheme for the Stochastic Simulation of Quantum Spin Dynamics
by Stefano De Nicola
This Submission thread is now published as
Submission summary
Authors (as registered SciPost users):  Stefano De Nicola 
Submission information  

Preprint Link:  https://arxiv.org/abs/2103.16468v2 (pdf) 
Date accepted:  20210716 
Date submitted:  20210701 19:53 
Submitted by:  De Nicola, Stefano 
Submitted to:  SciPost Physics 
Ontological classification  

Academic field:  Physics 
Specialties: 

Abstract
The numerical simulation of dynamical phenomena in interacting quantum systems is a notoriously hard problem. Although a number of promising numerical methods exist, they often have limited applicability due to the growth of entanglement or the presence of the socalled sign problem. In this work, we develop an importance sampling scheme for the simulation of quantum spin dynamics, building on a recent approach mapping quantum spin systems to classical stochastic processes. The importance sampling scheme is based on identifying the classical trajectory that yields the largest contribution to a given quantum observable. An exact transformation is then carried out to preferentially sample trajectories that are close to the dominant one. We demonstrate that this approach is capable of reducing the temporal growth of fluctuations in the stochastic quantities, thus extending the range of accessible times and system sizes compared to direct sampling. We discuss advantages and limitations of the proposed approach, outlining directions for further developments.
Author comments upon resubmission
Dear Editor and Referees,
I would like to thank you for your careful analysis and positive assessment of the present manuscript and for the Referees’ helpful and supportive comments. Answers to the Referees’ questions and comments are provided below and in the revised manuscript, as indicated in the list of changes.
Reply to Referee 1:
I would like to thank the Referee for their attentive reading of the manuscript and their very positive and supportive comments. Below I answer the Referee’s question.
 The Referee asks whether the exponential growth of fluctuations in the stochastic quantities can be interpreted in terms of a Lyapunov exponent (LE). To the best of my knowledge, LEs are typically defined for differential equations in order to quantify the divergence of solutions upon a small perturbation of the initial state, i.e. for trajectories ( x ) one has a LE ( \lambda ) when ( \delta x(t)  \approx e^{\lambda t} \delta x_0 ) . A positive maximal LE then signals (deterministic) chaotic behavior. In the present context, in contrast, all trajectories are initialized with identical initial conditions for all realizations of the SDEs, ( \delta \xi^a_i(0) =0 ), and the trajectories diverge only as a consequence of the subsequent stochastic evolution. Therefore, in the present case a LE in the above sense cannot be defined; however, it might be possible to quantify the degree of divergence in the stochastic trajectories by a suitable generalization of this quantity. I would like to thank the Referee for suggesting this connection, which could be an interesting direction for future work.
Reply to Referee 2:
I would like to thank the Referee for their attentive reading of the manuscript and for providing valuable inputs for clarification and further improvement. Below I answer the Referee’s questions and comments.
1 & 2. The dynamic structure factor can in principle be computed similarly to how it is done from other numerical techniques, e.g. tensor network methods, by first computing the required 2point functions and then numerically performing the required time integral and spatial sum, see e.g.: S. R. White and I. Affleck, Phys. Rev. B 77, 134437 (2008). Nonetheless, in practice only short times, corresponding to high frequencies, can presently be accessed; thus, an extrapolation method would currently be needed, the development of which is beyond the scope of the present work. However, it is worth noticing that this shortcoming is shared with the majority of numerical methods, including tensornetwork based approaches, which are similarly limited to short times for 2D systems. In spite of this limitation, several benchmarking possibilities are however still available for the approach. These include comparisons to exact diagonalization, as presently done in the manuscript, which gives access to arbitrarily late times for systems of up to 2025 spins, and comparisons to other numerical methods that are currently being developed, such as those of Refs [1720]. Recent developments in the field of ultracold atoms and trapped ions (see e.g. Refs [12]) could also make it possible to benchmark the importance sampling approach against experimental results by directly comparing quantities such as local expectation values. The Berry connection can be computed by performing adiabatic time evolution and numerically differentiating the instantaneous eigenstates with respect to the evolution’s driving parameters; the Berry curvature then requires an additional differentiation. This is again possible in principle within the present approach. However, the computational implementation of this is beyond the scope of the present manuscript, which is concerned with the development of the importance sampling method and a number of immediate benchmarks. Nonetheless, I would like to thank the Referee for bringing attention to these important quantities, which in future works could provide additional benchmarks and open up new applications of the disentanglement approach. In contrast, the fidelity (i.e. return probability) is one of the simplest quantities to compute within the stochastic approach, and it corresponds to the quantity ( A(t)^2 ) considered in Section V. In general, although any timedependent quantity can be represented in this language, the stochastic approach is most suited to computing quantities that can be readily expressed in terms of the timeevolution operator, such as local expectation values. A comment clarifying this was added after Eq. (9).

For the transversefield Ising model, it can be shown analytically that DQPTs occur upon quenching across a QCP, as demonstrated in Ref. [32]. However, this correspondence is not general: examples are known of DQPTs occurring when quenching within a phase, as well as of the absence of DQPTs in spite of quenching across a QCP, see e.g.: S. Vajna and B. Dóra, Phys. Rev. B 89, 161105(R) (2014). Similarly, critical exponents relative to DQPTs can be defined, see Ref. [8], but they do not appear to have a general relation to equilibrium exponents.

The current approach does not make assumptions about the strength of interactions; it is generally applicable to Hamiltonians of the form (1). Based on the derivation given in the manuscript, the optimal trajectory will be given by a meanfield trajectory in the singlespin basis. This is ultimately due to the fact that the disentanglement approach represents the system as an ensemble of noninteracting spins under the action of external fields. So, although the importance sampling approach is more generally applicable, it can be expected to be most advantageous in cases where mean field provides a reasonably good approximation to the true quantum dynamics. This is because the sampling has to account for the part of the dynamics that differs from mean field (i.e., for entanglement on top of the mean field evolution). This is now further clarified at the end of Section IV. This also lies at the root of the better performance observed for weak fields, since this regime approaches a classical limit, as remarked by the Referee in their minor comment 4; a comment clarifying this point was added at the end of Section IV. It is however possible that, for systems where a mean field approach would not work well, a similar importance sampling scheme could be formulated using a basis other than the single spin one. This would be an interesting direction for future developments, and I would like to thank the Referee for highlighting this point.
Reply to the minor comments:

The relation between ( \varphi ) and ( \phi ) is given in the paragraph following Eq. (4). Namely, they are related by a linear transformation given by the matrix ( O ), which is determined by the interaction matrix ( \mathcal{J} ). Clarifications on this were added in the paragraphs following Eqs (4) and (11).

In the present manuscript, for simplicity I focused on the timeevolution of the ( \downarrow ) spin coherent state. However, evolving any other spin coherent state poses no additional difficulty and can be done within the same formalism by simply adjusting the initial conditions ( \xi^a_i(0) ). This was previously shown in e.g. Ref. [25] or Ref. [31] for imaginary time evolution.

This point was addressed within item 4 of the main comments.

The stochastic approach can treat odd and even system sizes on an equal footing, see e.g. Ref. [23] where an Ising chain of 14 spins is considered. The importance sampling scheme does not add any additional complication relative to this. However, for the specific case of the Ising model additional care is needed as the interaction matrix is not invertible if any of the dimensions of the system is a multiple of 4. This is discussed in Ref. [24], where it is shown that this issue can be readily circumvented by introducing a diagonal shift to the interaction matrix ( \mathcal{J} ), which does not affect the ensuing dynamics. In the updated manuscript, a footnote was dedicated to a more detailed discussion of the explicit form of the interaction matrix ( \mathcal{J} ) for the Ising model, including a comment regarding sizes multiple of 4. A typo was also corrected. In the general derivations of the manuscript, no boundary conditions are assumed and the interaction matrix ( \mathcal{J} ) is allowed to be general. A comment clarifying this was added following Eq. (1). However, for numerical calculations we consider the Ising model (13) with periodic boundary conditions, as mentioned following Eq. (13).
5 & 6. I thank the Referee for suggesting these improvements; in the revised manuscript the missing physical parameters are now also provided in the main text of Sections IV and V, and the nomenclature has been uniformly changed to real/imaginary time.
List of changes
 Section I, after Eq. (1): added an explicit mention that no boundary conditions are assumed.
 Section II, after Eq. (4): added a comment on the definition of the \( O \) matrix.
 Section II, after Eq. (9): added a remark on which quantities can be expressed in the disentanglement formalism.
 Section III, after Eq. (11): added a comment pointing to the relevant discussion of the relation between \( \phi \) and \( \varphi \) fields.
 Section III, after Eq. (13): explicitly defined Ddimensional spatial indices to clarify the notation and facilitate the discussion of the interaction matrix.
 Section III, after Eq. (13): moved discussion of the explicit form of the matrix \( \mathcal{J} \) for the Ising model to a footnote; extended this discussion, corrected a previous typo and included a mention of the case of dimensions multiple of 4.
 Sections IV (bottomleft of p. 5) & V (topright of p. 6): added to the main text the missing physical parameters relative to figures.
 End of Section IV: clarified that for small \(\Gamma \) a classical limit is approached.
 End of Section IV: added comments on the range of applicability of the approach and the regimes where it can be expected to work best.
 Throughout the manuscript: changed the nomenclature “Euclidean time” to “imaginary time”.
Published as SciPost Phys. 11, 048 (2021)
Reports on this Submission
Report
The author has satisfactorily addressed my last remaining question, which has anyway been more related to my own scientific curiosity than targeting the quality of the manuscript. Along the lines of my previous report I very much recommend publication of the paper.