SBo=∫β0dtTr[12˙Φ2+12BΦ2].
In statistical mechanics (quantum field theory) we calculate the partition function (Euclidean path integral) given by
Z=∑ne−βEn=∫DΦexp(−SBo[Φ]ℏ).
This corresponds to the canonical ensemble where the temperature β≡1/ℏ is kept fixed. Here we usually employ the Metropolis algorithm, which is the most general of all Monte Carlo algorithms, to calculate this partition function. This is given explicitly by the probability distribution
P(Φn⟶Φ′n)=min(1,exp(−ΔSBo)) , ΔSBo=SBo[Φ′]−SBo[Φ].
In more detail, the Metropolis algorithm consists of the following steps:
- We start from some initial configuration Φ0.
- We propose a new configuration ˜Φ.
- We compute the variation of the action ΔSBo=SBo[Φ]−SBo[Φ0].
- We accept the new proposal with a probability P given by (3).
- We repeat starting from step 2 until we reach thermalization.
- We measure thermalized configurations which are used to calculate expectation values of physical observables.
The meaning of equation (3) is as follows. After we compute the variation of the action in step 3 we check its sign. If ΔSBo<0 then we should accept the proposal since it had resulted in a decrease of the action and thus getting us closer to the minimum. Otherwise, if ΔSBo>0 we accept the proposal only with a probability given by the Boltzmann distribution exp(−ΔSBo). This is the part which is simulating quantum mechanics and it is implemented numerically via the Von Neumann method. In other words, choose a uniform random number r between 0 and 1 and compare it to the Boltzmann distribution P=exp(−ΔS). If r<P then accept the proposal otherwise reject the proposal.
In many circumstances such as non-local theories it is found that the micro-canonical ensemble is much more favorable than the canonical ensemble. In this case it is the energy that is kept fixed and the preferred algorithm in this case is the hybrid Monte Carlo algorithm which synthesizes together the Metropolis algorithm, the molecular dynamics algorithm and the heat bath algorithm. This alternative formulation relies on the partition function
Z=∫DΦDPexp(−12Λ∑n=1TrP2n−SBo[Φ]ℏ).
The Euclidean action SBo acts therefore as a potential term while the new fields Pn act as conjugate momenta associated with the generalized coordinates Φn, i.e. ∑nTrP2n/2 is a kinetic energy. The Hamiltonian is then given by
HBo=12Λ∑n=1TrP2n+SBo[Φ]ℏ.
This Hamiltonian defines a canonical evolution in a fictitious time denoted by τ. Indeed, Hamilton equations of motion read explicitly
d(Pn)abdτ=−∂HBo∂(Φn)ba⇔−d(Pn)abdτ=∂SBo∂(Φn)ba≡(Fn)abd(Φn)abdτ=∂HBo∂(Pn)ba⇔d(Φn)abdτ=(Pn)ab.
The force Fn is given explicitly by
These equations are solved using a molecular dynamics algorithm. In particular, the so-called leap-frog algorithm, which preserves the phase space volume and reversibility and only break Hamiltonian conservation, gives us explicitly the equations
(Pn)ab(τ+δτ2)=(Pn)ab(τ)−δτ2(Fn)ab(τ).
(Φn)ab(τ+δτ)=(Φn)ab(τ)+δτ(Pn)ab(τ+δτ2).
(Pn)ab(τ+δτ)=(Pn)ab(τ+δτ2)−δτ2(Fn)ab(τ+δτ)..
We remark that the force Fn at the instant τ is needed to advance all the configurations Φn from τ to τ+δτ. Thus, we need to calculate the force Fn for all n at instant τ, then apply equations (7) and (8) to advance all Φn from τ to τ+δτ, after which we calculate again the force Fn for all n at instant τ+δτ, before we finally apply equation (9) to advance all Pn from τ+δτ/2 to τ+δτ.
The leap-frog will need two extra parameters: the step δτ and the number of iterations which we call L. The total time of the motion is given by T≡Lδτ. The initial values of Φn(0) and Pn(0) at time 0 will be specified and then by applying the above leap-frog algorithm we will obtain the final values of Φn(T) and Pn(T) at time T.
Two essential remarks can be stated now:
- The molecular dynamics involves in an obvious way a systematic error.
- The molecular dynamics probes only classical physics.
These two problems can be solved at once via the so-called hybrid Monte Carlo algorithm which is the most general of all Monte Carlo algorithms. This algorithm involves in an essential way the Metropolis algorithm. Indeed, the configuration Φn(T) obtained from the molecular dynamics algorithm is the solution which we will propose as a possible update Φ′n to the Metropolis algorithm (3). The probability distribution in this case becomes
P(Φn⟶Φ′n,Pn⟶P′n)=min(1,exp(−ΔHBo)) , ΔHBo=HBo[Φ′,P′]−HBo[Φ,P].
As it turns out, the conjugate momenta Pn should be updated using a heat bath algorithm in order to avoid ergodic problems, i.e. to be able to reach every point in phase space. This means in particular that Pn should be updated directly from a Gaussian distribution. This is indeed possible since the path integral over the conjugate momenta Pn is only Gaussian. In fact, this path integral can be given by a closed-form expression of the form
Z=∫DPexp(−12Λ∑n=1TrP2n)=(∫∏idPii∏i>jdPijdP∗ijexp(−12N∑i=1P2ii−∑i>jPijP∗ij))Λ.
All these integrals are Gaussian of the form
aπ∫dzdz∗exp(−azz∗)=∫10dv1∫10dv2.
z=reiθ , v1=exp(−ar2) , v2=θ2π.
These two equations show that v1 and v2 are uniform random numbers between 0 and 1 and thus z is a complex random number given by the formula
z=√−1alnv1(cos2πv2+isin2πv2).
The components of the conjugate momenta (Pn)ij are given by Gaussian random numbers of this form.
The action S=bTrP2n at each lattice point can also be rewritten as an eigenvalue problem. As it turns out, the eigenvalue distribution is given by the Wigner semi-circle law
ρ(x)=2πδ2√δ2−x2 , δ2=2Nb.
In summary, the {hybrid Monte Carlo algorithm} is an algorithm in which two crucial extra steps are added to the Metropolis algorithm:
- The step number 2 in the Metropolis algorithm is implemented via the {\bf molecular dynamics algorithm}. In other words, the new configurations Φ′n, P′n are given by the solutions Φn(T), Pn(T) of the molecular dynamics problem with Φn(0), Pn(0) as the initial conditions.
- As it turns out the path integral over the conjugate momenta Pn should be sampled using the so-called {\bf heat bath algorithm} in order to avoid the ergodic problem, i.e. to be able to reach every point in phase space.
No comments:
Post a Comment