Concepts

Introduction & Motivations

  • Purpose:

    The purpose of this little class is too facilitate the writing and maintenance of Metropolis MonteCarlo algorithms, with several advantages :

    • Easy to add new moves and measures.

    • For complex MC algorithms, with 6 or 7 different moves, it helps to clearly separate the code for each move.

    • Parallelism is automatic.

    • The random generator is a simple parameter, it can be chosen dynamically.

  • Principle

    The mc_generic class is a generic version of the algorithms, with moves and measures. The user :

    • writes move classes, modelling the Move concept.

    • writes measure classes, modelling the Measure concepts.

    • register them in a mc_generic object MC

    • call MC.warmup_and_accumulate(…) … and that is (almost) it !

The Move concept

Elements

Comment

  • mc_sign_type attempt()

  • First part of the Move.

  • Returns the probability to accept the move. If :

    • the move is \(x\rightarrow x'\), proposed with proba \(T_{x\rightarrow x'}\)

    • the probability of the \(x\) config is denoted \(p_x\)

    • then attempt should return the Metropolis ratio:

    \[\frac{p_{x'} T_{x'\rightarrow x}}{p_x T_{x\rightarrow x'}}\]

    with the sign. (The sign will be separated from the absolute value and kept by the MonteCarlo class). In other words: attempt() = move_sign * move_rate with abs(move_sign) = 1

  • mc_sign_type accept()

  • Called iif the Move is accepted.

  • Returns a number r such that \(|r| =1\)

  • There is a possibility to update the sign here as well, so that the final sign is: move_sign * accept()

  • CAREFUL that we need the ratio new_sign / old_sign here just like we needed a ratio in attempt()

  • void reject()

  • Called iif the Move is rejected (for cleaning).

The Measure concept

Elements

Comment

  • void accumulate(std::complex<double> sign)

  • Accumulation with the sign

  • void collect_results ( boost::mpi::communicator const & c)

  • Collects the results over the communicator, and finalize the calculation (compute average, error).