Calculation of the Metropolis and the Glauber
Transition Probabilities for the Ising Model
and for the q-state Potts Model
by Peter Meyer
Appendix I from the author's M.Phil. thesis
Hermetic Systems

(i) Transition Probabilities in the Ising Model

Consider a spin model in a specific state, a particular spin Si and the set { Sr : Sr is a nearest neighbour of Si }. The energy Ei contributed by the spin Si is the sum of its interaction energies with the Sr and so

Ei = -J.Σr(Si.Sr) = -J.SirSr

In the Ising model there are only two possible spin values, +1 and -1, so the only possible change to the spin Si is for it to take the opposite value, namely, -Si. Call this spin Sj . In this case the energy Ej associated with the new spin value is the sum of the interaction energies of Sj with the Sr and so

Ej = -J.Σr(Sj.Sr) = -J.Σr(-Si.Sr) = -J.-SirSr = J.SirSr

Thus the difference in the "before" and "after" energies is

ΔE = Ej - Ei = ( J.SirSr ) - ( -J.SirSr ) = 2.J.SirSr                  (1)

Suppose that there is a maximum number of nearest neighbours which can be possessed by any spin, and let this number be denoted by n. Then the maximum value for ΣrSr is +n and the minimum value is -n. Thus there are, at most, 2n+1 possible values for ΔE, corresponding to the possible values for ΣrSr , namely:

-n, -n+1, -n+2, ..., -2, -1, 0, 1, 2, ..., n-1, n

although in a particular Ising spin model not all of these values may in fact be possible.

The transition probabilities for the Metropolis algorithm and for the Glauber algorithm were given in Section 1.7 as follows:

Metropolis algorithm: W(Si Sj)= 1 if ΔEji ≤ 0
 = eEji/(kBT) otherwise.
Glauber algorithm: W(Si Sj)= 1 / [ 1 + e ΔEji/(kBT) ]

where T is the temperature and kB is Boltzmann's constant. From equation (1) we obtain:

Metropolis: W(Si Sj)= 1 if 2.J.SirSr ≤ 0
 = e -2JSirSr/kBT otherwise.
Glauber: W(Si Sj)= 1 / [ 1 + e 2JSirSr/(kBT) ]

Taking J = kB = 1 the algorithms become:

Metropolis: W(Si Sj)= 1 if 2.SirSr ≤ 0
 = e -2.SirSr/T otherwise.
Glauber: W(Si Sj)= 1 / [ 1 + e 2.SirSr/T) ]

Let S2 = 2.SirSr then we have:

Metropolis: W(Si Sj)= 1 if S2 ≤ 0
 = e -S2/T otherwise.
Glauber: W(Si Sj)= 1 / ( 1 + eS2/T )

There are at most 2n+1 possible values for S2 (assuming a maximum of n nearest neighbours) as follows:

-2n, -2n+2, -2n+4, ..., -2, 0, 2, ..., 2n-2, 2n

so the transition probabilities that we seek are:

Metropolis:
S2:    -2n    -2n+2   ...    -2    0      2     4    ...     2n
W:      1       1     ...     1    1    e-2/T  e-4/T   ...    e-2n/T

Glauber:
S2:      -2n        ...        -2          0        2      ...      2n
W:    1/(1+e-2n/T)   ...    1/(1+e-2/T)    1/2    1/(1+e2/T)  ...    1/(1+e2n/T)

(ii) Transition Probabilities in the Potts Model, Version A.

Consider a spin model in a specific state, a particular spin Si (1 ≤ Si ≤ q) and the set { Sr : Sr is a nearest neighbour of Si }. The energy Ei associated with the spin Si is the sum of its interaction energies with the Sr and so Ei = -J.Σr[δ(Si,Sr)].

In the q-state Potts model there are q possible spin values, 1, ..., q, so the spin Si may change to any of q-1 possible new values. We select one at random; call this spin Sj. In this case the energy Ej associated with the new spin value is the sum of the interaction energies of Sj with the Sr and so Ej = -J.Σr[δ(Sj,Sr)], where δ() is the Kronecker delta function: δ(x,y) = 1 if and only if x = y (otherwise 0). Thus the transition energy, ΔE, is

Ej - Ei = -J.Σr[δ(Sj,Sr)] - -J.Σr[δ(Si,Sr)] = -J.{ Σr[δ(Sj,Sr)] - Σr[δ(Si,Sr)] } so

ΔE = J.{ Σr[δ(Si,Sr)] - Σr[δ(Sj,Sr)] }                  (2)

Suppose that there is a maximum number of nearest neighbours which can be possessed by any spin, and let this number be denoted by n. Then the maximum value for each of Σr[δ(Si,Sr)] and Σr[δ(Sj,Sr)] is +n and the minimum value is 0. Thus the maximum value for

Sr[δ(Si,Sr)] - Σr[δ(Sj,Sr)]                  (3)

is +n and the minimum is -n, so (as with the Ising model) there are, at most, 2n+1 possible values for ΔE (and thus for the transition probabilities) depending on the possible values for (3), namely:

-n, -n+1, -n+2, ..., -2, -1, 0, 1, 2, ..., n-1, n

although in a particular q-state Potts spin model not all of these values may in fact be possible.

The Metropolis and the Glauber algorithms for the q-state Potts model, version A, are obtained by substituting the RHS of (2) for ΔE into the definitions given in Section (i) above. As before we take J = kB = 1, so the algorithms become:

Metropolis: W(Si Sj) = 1 if Σr[δ(Si,Sr)] - Σr[δ(Sj,Sr)] ≤ 0,

= e-{Sr[δ(Si,Sr)] - Sr[δ(Sj,Sr)]}/T otherwise,

that is, W(Si Sj) = 1 if Σr[δ(Si,Sr)] ≤ Σr[δ(Sj,Sr)],

= e-{Sr[δ(Si,Sr)] - Sr[δ(Sj,Sr)]}/T otherwise.

Glauber: W(Si Sj) = 1 / ( 1 + e{Sr[δ(Si,Sr)] - Sr[δ(Sj,Sr)]}/T )

Let S3 = Σr[δ(Si,Sr)] - Σr[δ(Sj,Sr)] then we have:

Metropolis: W(Si Sj) = 1 if S3 ≤ 0,

= e -S3/T otherwise.

Glauber: W(Si Sj) = 1 / ( 1 + eS3/T )

There are at most 2n+1 possible values for S3 (assuming a maximum of n nearest neighbours) as follows:

-n, -n+1, -n+2, ..., -2, -1, 0, 1, 2, ..., n-1, n

so the transition probabilities that we seek are:

Metropolis:
S3:  -n    -n+1   ...    -1    0      1       2   ...     n
W:    1     1     ...     1    1    e-1/T    e-2/T   ...    e-n/T

Glauber:
S3:      -n       ...        -1         0        1      ...       n
W:    1/(1+e-n/T)  ...    1/(1+e-1/T)   1/2    1/(1+e1/T)  ...    1/(1+en/T)

It will be noted that these are not the same transition probabilities as in the Ising case.

(iii) Transition Probabilities in the Potts Model, Version B.

We follow the same reasoning as in the previous section.

Consider a spin model in a specific state, a particular spin Si (1 ≤ Si ≤ q) and the set { Sr : Sr is a nearest neighbour of Si }. The energy Ei associated with the spin Si is the sum of its interaction energies with the Sr and so Ei = -J.Σr[2.δ(Si,Sr)-1].

In the q-state Potts model there are q possible spin values, 1, ..., q, so the spin Si may change to any of q-1 possible new values. We select one at random; call this spin Sj. In this case the energy Ej associated with the new spin value is the sum of the interaction energies of Sj with the Sr, so Ej = -J.Σr[2.δ(Sj,Sr)-1]. Thus the transition energy, ΔE, is

Ej - Ei = -J.Σr[2.δ(Sj,Sr)-1] - -J.Σr[2.δ(Si,Sr)-1] = -J.{ Σr[2.δ(Sj,Sr)] - Σr[2.δ(Si,Sr)] } so

ΔE = 2.J.{ Σr[δ(Si,Sr)] - Σr[δ(Sj,Sr)] }                  (4)

Suppose that there is a maximum number of nearest neighbours which can be possessed by any spin, and let this number be denoted by n. Then the maximum value for each of Σr[δ(Si,Sr)] and Σr[δ(Sj,Sr)] is +n and the minimum value is 0. Thus the maximum value for

Sr[δ(Si,Sr)] - Σr[δ(Sj,Sr)]                  (5)

is +n and the minimum is -n, so (as with the Ising model and the Potts model, Version A) there are, at most, 2n+1 possible values for ΔE (and thus for the transition probabilities) depending on the possible values for (5), namely:

-n, -n+1, -n+2, ..., -2, -1, 0, 1, 2, ..., n-1, n

although in a particular q-state Potts spin model not all of these values may in fact be possible.

The Metropolis and the Glauber algorithms for the q-state Potts model, Version B, are obtained by substituting the RHS of (4) for ΔE into the definitions given in Section (i) above. As before we take J/kB = 1, so the algorithms become:

Metropolis: W(Si Sj) = 1 if 2.{ Σr[δ(Si,Sr)] - Σr[δ(Sj,Sr)] } ≤ 0,

= e-2.{Sr[δ(Si,Sr)] - Sr[δ(Sj,Sr)]}/T otherwise,

that is, W(Si Sj) = 1 if Σr[δ(Si,Sr)] ≤ Σr[δ(Sj,Sr)],

= e-2.{Sr[δ(Si,Sr)] - Sr[δ(Sj,Sr)]}/T otherwise.

Glauber: W(Si Sj) = 1 / ( 1 + e2.{Sr[δ(Si,Sr)] - Sr[δ(Sj,Sr)]}/T )

Let S4 = 2.{Σr[δ(Si,Sr)] - Σr[δ(Sj,Sr)]} then we have:

Metropolis: W(Si Sj) = 1 if S4 ≤ 0,

= e -S4/T otherwise.

Glauber: W(Si Sj) = 1 / ( 1 + eS4/T )

There are at most 2n+1 possible values for S4 (assuming a maximum of n nearest neighbours) as follows:

-2n, -2n+2, -2n+4, ..., -2, 0, 2, ..., 2n-2, 2n

so the transition probabilities that we seek are:

Metropolis:
S4:   -2n   -2n+2    ...    -2    0      2       4    ...    2n
W:     1       1     ...     1    1    e-2/T    e-4/T    ...    e-2n/T

Glauber:
S4:      -2n         ...        -2         0         2       ...       2n
W:    1/(1+e-2n/T)    ...    1/(1+e-2/T)    1/2   1/(1+e2/T)    ...    1/(1+e2n/T)

It will be noted that these are not the same transition probabilities as for Version A of the Potts Model, but they are the same as those for the Ising model.

(iv) The Equivalence of the 2-state Potts Model and the Ising Model.

It is commonly said that the 2-state Potts model is equivalent to the Ising model. This is true if Version B of the Potts model is meant, because then the transition probabilities are the same and so the dynamics of the two models are the same. Usually, however, when the Potts model is discussed it is Version A which is meant, and in this case, even for q=2, the transition probabilities are not the same. Thus at the microscopic level there are differences between the Ising model and the 2-state Potts model, Version A, so we are led to ask whether the 2-state Potts model, Version A, is strictly equivalent to the Ising model.

The transition probabilities (in both the Metropolis algorithm and the Glauber algorithm) are determined by (i) the transition energies, i.e., the energy difference between the initial state and a possible new state, (ii) the value of J/kB and (iii) the temperature. Since (ii) and (iii) are not properties of the spin model itself the question becomes whether (i) is the same regardless of whether the spins are described in terms of one model definition (Ising) or in terms of the other (2-state Potts, Version A). We shall first consider the 2-state Potts model, Version B.

Version B:

Consider a spin Si and the set S = { Sr : Sr is a nearest neighbour of Si }. S may be partitioned into S1 = { Sr : Sr = Si } and S2 = { Sr : Sr ≠ Si }.

Suppose Si and the Sr are viewed as Ising spins. Suppose Si = +1, then (from section (i) above and the definition of the Ising model in Section 1.5):

ΔEIsing = 2.JIsing.(+1).ΣrSr = 2.JIsing.( |S1| - |S2| )

Suppose Si = -1, then

ΔEIsing = 2.JIsing.(-1).ΣrSr = -2.JIsing.( -|S1| + |S2| ) = 2.JIsing.( |S1| - |S2| )

Thus in each case ΔEIsing = 2.JIsing.( |S1| - |S2| ).

Now suppose Si and the Sr are viewed as spins in the 2-state Potts model, Version B. Then (from the definition of the Potts model, Version B, in Section 1.5)

ΔEPotts,B = -JPotts,B.2.{ Σr[δ(Sj,Sr)] - Σr[δ(Si,Sr)] }

where Sj is some spin value other than Si, so

ΔEPotts,B = -JPotts,B.2.( |S2| - |S1| ) = JPotts,B.2.( |S1| - |S2| )

Since ΔEIsing = ΔEPotts,B we may conclude that the Ising model is strictly equivalent to the 2-state Potts model, Version B.

Version A:

As above, consider a spin Si and the set S = { Sr : Sr is a nearest neighbour of Si }. S may be partitioned into S1 = { Sr : Sr = Si } and S2 = { Sr : Sr ≠ Si }. As before, ΔEIsing = 2.JIsing.( |S1| - |S2| ).

Now suppose Si and the Sr are viewed as spins in the 2-state Potts model, Version A. Then (from the definition of the Potts model, Version A, in Section 1.5)

ΔEPotts,A = -JPotts,A.{ Σr[δ(Sj,Sr)] - Σr[δ(Si,Sr)] }

where Sj is some spin value other than Si, so

ΔEPotts,A = -JPotts,A.( |S2| - |S1| ) = JPotts,A.( |S1| - |S2| )

Since ΔEIsing = 2.ΔEPotts,A the transition energies (and thus the transition probabilities) are different.

Okano et al. (1997, p.738) state: "It is known that the critical points [of the q-state Potts model] locate at Jc = log(1+√q)." Taking Tc = 1/Jc and q = 2 we obtain

Tc = 1/ln(1+√2) = 1.134593

This is exactly one-half of the value usually given for the Ising model, namely, 2/ln(1+√2) = 2.269185 (see e.g. Stinchcombe 1983, p.177, taking J = 1 = kB), which value is obtained from simulations. Thus although the Ising model and the 2-state Potts model are commonly said to be equivalent, the values of the critical temperatures are not the same when Version A of the 2-state Potts model is used.

Nevertheless, the critical exponents of the Ising model and the 2-state Potts model, Version A, are found to be the same, the only difference being the value of the critical temperature. Thus the difference between the two models is really only one of the scale used for the temperature.

Consider the Potts model, Version A, in which the interaction energy is JPotts,A = 2.JIsing then

ΔEPotts,A = 2.JIsing.( |S2| - |S1| ) = ΔEIsing

so we can also say that the 2-state Potts model, Version A, is strictly equivalent to the Ising model provided the interaction energy of the 2-state Potts model is double that of the Ising model. Since the units of the critical temperature are related to units of J, and a J-unit in this modified Potts model, Version A, denotes twice the energy of a J-unit in the Ising model, the numerical value of the critical temperature in the former model is thus one-half of its value in the latter model, thus explaining the difference in the numerical values of the critical temperature in the Ising model and in the 2-state Potts model, Version A.

Computational Studies of Pure and Dilute Spin Models
Computational Science Hermetic SystemsHome Page