10

My question is about extracting observables from QMC methods, as described in this reference.

I understand the formal derivation of various QMC methods like Path Integral Monte Carlo. However, at the end of the day I'm still confused about on how to effectively use these techniques.

The basic idea of the derivation of Quantum MC methods is to discretise, via the Trotter approximation, an operator which may be either the density matrix or the time-evolution operator of a quantum system. We then obtain a classical system with an additional dimension which may be treated with MC methods.

Given that we can interpret $\beta$ in the quantum operator $e^{−\beta\hat{H}}$ both as an inverse temperature and an imaginary time, the aim of these algorithms should be to compute an approximation of this operator. Indeed, if we would directly measure quantities from the various configurations sampled along a simulation, in the "inverse temperature" case we would have samples respecting a probability density based on $\beta/M$, where $M$ is the number of discrete steps introduced in the Trotter decomposition. Instead, in the "imaginary time" case we would obtain samples at various discrete time-steps, thus getting averages along the time as well. We also wouldn't obtain quantities like $\langle\psi_t|\hat{A}|\psi_t\rangle$ at a given time $t$, with $\hat{A}$ some observable operator.

However, in my opinion the quantities we sample directly from this kind of simulations (taken from (5.34) of the document, page 35):

$$\bar{O} \equiv \langle \hat{O}(X) \rangle \equiv \frac{1}{N!} \sum_P \int O(X) \pi(X,P) dX $$

cannot be quantities related to the quantum system, given the additional dimension. Instead, the correct quantum quantities can be computed via formulas like (5.35), which contains in every sample a whole chain of $M$ simulated configurations:

$$\frac{E_{th}}{N}= \left\langle \frac{d}{2 \tau} - \frac{m}{2 (\hbar \tau)^2MN } \sum_{j=1}^{M} (\mathbf{R}_j -\mathbf{R}_{j+1})^2 + \frac{1}{MN} \sum_{j=1}^{M} V(\mathbf{R}_j) \right\rangle $$

Am I right that a series of QMC simulations is required to extract useful information about a given observable?

Dan
  • 3,355
  • 3
  • 21
  • 47
Pippo
  • 119
  • 6
  • Provided that I understand you correctly, it strikes me that the two approaches are equivalent if the system is ergodic. – Daniel Shapero Jul 23 '13 at 16:03
  • @DanielShapero What do you mean exactly with being equivalent? – Pippo Jul 24 '13 at 12:02
  • I just googled path integral Monte Carlo and you should actually just ignore what I said, it's irrelevant. – Daniel Shapero Jul 24 '13 at 18:16
  • I don't think there is any doubt regarding Quantum Monte Carlo; it's very well understood and rigorously backed theoretically... – Nick Aug 01 '13 at 16:11
  • What do you mean by $\beta/M$? $\beta$ is a number and if $M$ is a discretization like you said, it's a set. Is $M$ intended to be the trotter number? – Dan Aug 01 '13 at 20:04
  • "Given that we can interpret β in the quantum operator e−βH^ both as an inverse temperature and an imaginary time, the aim of these algorithms should be to compute an approximation of this operator. " I don't see how that goal follows from that property of the $e^{-\beta \hat{H}}$ operator. – Dan Aug 01 '13 at 20:22
  • @Pippo: You might want to elaborate in your question. It's difficult to figure out what exactly your question is asking. In particular, it's difficult to tell what you think the aim of these algorithms is given what you say you think it should be. – Dan Aug 01 '13 at 20:50
  • @Dan: Yes, $M$ is the Trotter number; I will give more details of waht I would like to know as comments to your answer. – Pippo Aug 02 '13 at 09:09
  • For further clarification, by "the additional dimension" do you mean $\beta$? – Dan Aug 03 '13 at 08:06
  • @Dan: Yes, exactly; or better, the discretized version of $\beta$, i.e. $\beta/M$ introduced by the Trotter approximation. – Pippo Aug 04 '13 at 11:42

2 Answers2

2

There is a lot of confusion in your question. The most important for me is that you miss that "naive" QMC which is Monte-Carlo calculation of integrals in some variational method and diffusion Monte-Carlo are different methods with different argumentation and derivation.

The main point however is about imaginary time. In diffusion Monte-Carlo imaginary time is a trick to convert time-independent Schroedinger equation into time-dependent diffusion-like equation which solution in the infinite "time" limit tends to a solution of original Schroedinger equation. That's it. Time in DQMC is not real.

Relatively good but simple explanation is given in Reviews of Modern Physics, 73, 33(2001).

P.S. By the way, what you mean by "Trotter approximation" in your question?

Misha
  • 240
  • 1
  • 6
  • I don't think there is this confusion is my question since I never referred to Diffusion MC, whose idea is quite different, although it also starts from a discretization of the density/time-evolution operator (but it ends with a different interpretation of it). – Pippo Oct 06 '13 at 10:35
  • With "Trotter approximation" I mean the idea of approximating the operator $e^{−\beta\hat{H}}$ with the product $e^{−\tau\hat{H}}\cdot e^{−\tau\hat{H}}\cdot ... \cdot e^{−\tau\hat{H}}$, with $\tau$ being $\beta/M$. – Pippo Oct 06 '13 at 10:37
  • Btw, at the end I solved my issue asking directly to the professor at the end of the exam (which went very well :D ), and yes, we can't direclty link simulated quantities to the quantum desired ones. – Pippo Oct 06 '13 at 10:40
  • @Pippo So, what you actually meant was Path-Integral Monte Carlo. I still do not see you mention this in your question. – Misha Oct 06 '13 at 14:34
  • Second line: "I understand the formal derivation of various QMC methods like Path Integral Monte Carlo." ;) – Pippo Oct 09 '13 at 07:58
  • Yes, you mention it in your question. Your question is formulated as a general for various QMC, which is misleading. – Misha Oct 09 '13 at 10:18
0

You're right that people use monte carlo techniques to calculate statistical averages (as opposed to time-resolved information) all the time. It's not necessarily true that this is what should be calculated: it depends on what kind of information you want. Maybe you have a time-dependent external forcing, for example, and want to see how the system evolves in response.

Dan
  • 3,355
  • 3
  • 21
  • 47
  • Thank you for answering. I will try to give more details of what I am asking. To make myself more understandable, I will refer to this work which I found on the Internet: http://www.itp.phys.ethz.ch/education/fs12/cqp/chapter05.pdf – Pippo Aug 02 '13 at 09:11
  • The basic idea of the derivation of Quantum MC methods is to discretise, via Trotter approximation, an operator which may be either the density matrix or the time-evolution operator of a quantum system; in this way, we obtain a classical system with an additional dimension which may be treated with MC methods. – Pippo Aug 02 '13 at 09:12
  • However, in my opinion the quantities we sample directly from this kind of simulations (see (5.34) of the document, page 35) cannot be quantities related to the quantum system, given the additional dimension. Instead, the correct quantum quantities can be computed via formulas like (5.35), which indeed contains in every "sample" a whole chain of $M$ configurations. – Pippo Aug 02 '13 at 09:16
  • Here comes my question: am I right with my interpretation of Quantum Monte Carlo methods? – Pippo Aug 02 '13 at 09:17
  • I will also edit my original question. – Pippo Aug 02 '13 at 09:18