Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.

html

Random physics (/)

Alberto Verga, research notebook

blog (/) | about (/pages/about.html)

lectures (/pages/lectures.html)

Lectures (/pages/PS-index.html) on statistical mechanics.

Onsager phase transition: the square

lattice ferromagnet
Following an original proposition of Pierre Weiss, Lenz around 1920 thought that a magnetic material
was formed by elementary interacting magnets, however, at variance to Langevin and later Weiss, he
assumed discrete and not continuous values of the magnetic moment or “magneton” (Weiss, 1911) 1.
Ising introduced in his thesis (1924)2 the idea of nearest neighbor interaction between “spins” taking
values s = ±1. The ising hamiltonian in a twodimensional lattice reads,

∑ ∑ ∑
H = −Jx sxy sx+1y − Jy sxy sxy+1 − B sxy ,
x,y s x,y

where x = 0, 1, … , L − 1 and y = 0, 1, … , M − 1 are the coordinates of the lattice sites (the length
unit is the lattice step a = 1). The exchange energies in the x and y directions are Jx and Jy , B is the
external applied field (in energy units), and N = LM is the number of sites. The onedimensional case
corresponds to L = 1 (or M = 1). The configuration space is generated by the 2N vectors of the form,

s ∈ S = {(s0 , … , sn , … , sN−1 ), sn = ±1, n = 0, … , N − 1} .

Before considering the planar case, we rework the linear ising model in a way such that we can
generalize the transfer matrix method. We write the onedimensional hamiltonian as,

∑ ∑
H1D = −J (sx sx+1 − 1) − B sx
x x
J B
(sx − sx+1 )2 −
2∑ ∑ x
= (s + sx+1 )
x
2 x

(J is the coupling constant in 1D) which gives a ground state at zero energy (for B = 0 ). The second
form is symmetric in sx , sx+1 . Using the first equation we can write the partition function in terms of a
product of 2 × 2 matrices,

Z(K, h) = Tr e−H1D /T = Tr(V1 V3 )N

or using the cyclic property of the trace, we can obtain the symmetric form,

1 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

Z(K, h) = Tr (V31/2 V1 V31/2 ) = Tr V N ,

N

where V = V31/2 V1 V31/2 is the so called transfer matrix,3 and

( e−2K 1 )
1 e−2K
V1 (K) = ,

and

(0 e−h )
eh 0
V3 (h) = ,

with K = J/T and h = B/T . Using instead the symmetric form (the transfer matrix V ),

( e−2K e−h )
eh e−2K
V= ,

we get a simpler expression to compute Z . Indeed, for the symmetric form of the transfer matrix the
partition function can be evaluated using its eigenvalues λ± ,

Z(K, h) = Tr V N = λN+ + λN− ≈ λN+ , λ+ > λ−

where the last approximation is valid in the thermodynamic limit N → ∞. This is not possible in the
other case, where V is expressed as a product of two noncommuting matrices.

Instead of directly computing the eigenvalues of V , we observe that for twodimensional matrices we
have the identity EX,
̂
eεn⋅σ = cosh(ε) 12 + sinh(ε) (n̂ ⋅ σ)

which will allow us to transform V1 , V3 and V into the exponentials of a matrix; in the previous
formula ε is a constant, n̂ a unit vector, 12 the unit matrix, and σ the vector of pauli matrices.
Identifying the exponential form of V1 and V3 is straightforward:

V3 = ehσ3

and
⋆
−2K eK σ1
V1 = 1 + e σ1 =
cosh(K ⋆ )

where we defined the new constant,

tanh(K ⋆ ) = e−2K .

In the case of the matrix

‾‾‾‾V‾e−H ,
V = √det H = εn̂ ⋅ σ .

we should solve the system of equations in ε and n̂ , to obtain EX,

2 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

cosh ε = cosh K ⋆ cosh h

(sinh K ⋆ , 0, cosh K ⋆ sinh h)
n̂ = .
√‾cosh
‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾
K ⋆ cosh h − 1‾

The eigenvalues of H = εn̂ ⋅ σ are readily obtained,

ε = ± ln[cosh K ⋆ cosh h + √‾cosh K cosh2 h − 1‾]

‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾
2 ⋆

(the eigenvalues of n̂ ⋅ σ are obviously ±1).

It is interesting to note that the relation between K and K ⋆ is dual, in the sense that EX,
⋆
tanh(K ⋆ ) = e−2K , tanh(K) = e−2K .

EX Compute the free energy per spin and show that the mean magnetization is given by

sinh h
⟨s⟩ =
√‾cosh
‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾
K cosh2 h − 1‾
2 ⋆

and show that this result is equivalent to the one of section Lattices (/pages/PS-lattice.html).

The alternative formulation in terms of the product V1 V3 , can be handled in a particular limit, which
will allow us to conveniently combine the two exponentials. The interest of the method is that it can be
generalized to our case of interest, the twodimensional lattice. The Baker-Campbell-Hausdorff formula
says, for two operators operators A, B ,
2
eτA eτB = eτ(A+B)+(τ /2)[A,B]+⋯

where τ is a real constant, and the omitted terms are higher order in τ. In the limit τ → 0 one may
replace the product of exponentials by the exponential of the sum. We define new parameters λ and τ,
such that,

K ⋆ = λτ, h=τ

and write,

1
V1 V3 ≈ e−τH , Hλ = −λσ1 − σ3 .
cosh K ⋆

(Note that the prefactor in this expression is independent of h , thus irrelevant for the computation of the
magnetic properties of the system.) Therefore, in the low energy limit the onedimensional ising model
can be described by a hamiltonian Hλ corresponding to a single quantum spin. In this limit τ ∼ 1/T
becomes a continuous imaginary time of a quantum system: τ = it.4 One may establish the analogy:

Quantum system Statistical system

D dimensions D + 1 dimensions

3 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

Quantum system Statistical system

ground state equilibrium state

expected value ensemble average

fundamental energy free energy

The eigenvalues of τHλ are,

ε± = ±τ√‾λ‾‾‾‾
2
+ 1‾

which gives a free energy (per spin)

f (λ) = −T ε− = −Tτ√‾λ‾‾‾‾
2
+ 1‾,

remark that only the lowest eigenvalue (fundamental level of Hλ ) contributes to the free energy in the
thermodynamic limit, which gives in turn, the mean magnetization per spin,

∂f h
⟨s⟩ = − = .
∂h √‾K‾‾‾‾‾‾‾
⋆2 + h‾
2

Now we may compare the result of the exact method with the continuum limit result. In the limit of
small K ⋆ and h , the exact eigenvalue ε can be approximated by:

ε = ln[cosh K ⋆ cosh h + √‾cosh K cosh2 h − 1‾] ≈ √‾K‾‾‾‾‾‾‾

‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾
2 ⋆ ⋆2
+ h‾2 , (K ∗ , h → 0),

in accordance with the Hλ result EX.

We apply now the continuum limit method used to obtain Hλ , to the twodimensional lattice case. We
consider the B = 0 case, which is enough to exhibit the existence of a phase transition, our main goal,
through the existence of a singularity in the free energy for some value of the parameters T, Jx , Jy . Our
first task is to find the transfer matrix. We then plan to transform the transfer matrix into a unique
exponential of an effective hamiltonian easily amenable to a diagonal form.

4 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

Twodimensional square lattice in the (x, τ) coordinates; the x axis remains discrete while the τ
axis goes to the continuum limit.

In analogy with the onedimensional case we will treat the y direction as a (imaginary) time, and the
other direction as a normal discrete linear lattice; the transfer matrix will then represent the evolution
operator of the discrete system in τ (see the above figure). The physical properties of the system depend
on two parameters:

Kx = Jx /T, Kτ = Jy /T .

With this notation the partition function is,

L−1 L−1

[ ∑ ] [ ∑∑( x ]
s (τ)sx (τ + 1) − 1) ,
∑ ∑ x
Z= exp Kx s (τ)sx+1 (τ) exp Kτ
s τ x=0 τ x=0

where we substracted a constat energy term in H . The first exponential is diagonal in the variable τ, and
can be compared with the h term (and the matrix V3 ) in the onedimensional case, while the second
exponential represents the interaction between successive rows, and can be treated in the variable τ as
the onedimensional system in the variable x (and the matrix V1 ): instead of summing on the two
orientations of the spin, we sum over the whole configuration of the row. Therefore, we can write:

Z = Tr(V1 V3 )M

with
⋆
eKτ ∑x σ1 (x)
V1 =
(cosh Kτ⋆ )L

and

V3 = eKx ∑x σ3 (x)σ3 (x+1) .

At variance to the onedimensional lattice, the V matrices are 2L × 2L which makes the search for the
eigenvalues much more difficult.

The exact computation of the square lattice system partition function was obtained in 1944 by Onsager. 5

5 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

Later, in 1964, Schultz, Mattis and Lieb found a mapping to a fermion problem that allowed them to
greatly simplify the original Onsager’s computation.6 Here, we are interested in the thermodynamic
limit, in which the system is translationally invariant and independent of the boundary conditions. In
this limit the computation of the free energy using the fermion method is straightforward. However,
before discussing the continuous limit, let us explore the parameter phase space (Kx , Kτ ).

Parameter phase space of the anisotropic 2D ising model. The line tanh Kx = e−2Kτ separates
the disorderd (paramagnetic), when both Kx , Kτ are small, and ordered (ferromagnatic) phases,
for large coupling. The isotropic case corresponds to the line Kx = Kτ ; the continuous limit is
on the small Kx = τ side: it contains a transition point (black dot). A stretching of the vertical
axis allows to recover the isotropic case (black arrow) Kc = −0.5 ln(√2 ‾ − 1) .

In the isotropic case we have Kx = Kτ and the duality relation:

tanh K ⋆ = e−2K , or sinh 2K ⋆ sinh 2K = 1 .

This last relation shows that when K is large (at fixed J , this means low temperature), then K ⋆ must be
large (high temperature), and vice versa. The common region where both low and high temperature
properties overlap, is the transition region, corresponding then to the fixed point

K = K ⋆ = Kc , tanh Kc = e−2Kc

which defines the critical temperature Tc = −2/ ln(√2 ‾ − 1) = 2.269 J : below Tc (small K ⋆ ) the
system is ordered and disordered above it (large K ⋆ ).

In the anistropic case we have the duality relations:

tanh Kτ⋆ = e−2Kτ , tanh Kτ⋆ = e−2Kτ

and, a point in the plane (Kx , Kτ ), corresponds to a point in the dual plane (Kτ⋆ , Kx⋆ ) (note the inversion
of indices in the dual plane); the self-dual points are then on the line

(Kx , Kτ ) = (Kτ⋆ , Kx⋆ )

6 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

contains the critical points:

sinh 2Kx sinh 2Kτ = 1

represented in the figure to separate the two phases.

Therefore, using the parametrization,

Kτ⋆ = λτ, Kx = τ

we know that for some value of λ, in fact λc = 1 as we shall demonstrate, there is a phase transition, we
can combine the two exponentials in the small τ limit.This parametrization rescale the τ axis to recover,
in the continuum limit, the isotropic transition (corresponding to the arrow mapping the anisotropic
critical point to the isotropic one, in the figure above). Consequently, from the expressions of V1 and
V3 , in the small τ limit we write:

e−τH(λ)
∑ ∑
Z(λ) = Tr , H(λ) = −λ σ1 (x) − σ3 (x)σ3 (x + 1) . (1)
(cosh Kτ⋆ )L x x

When λ → 0 we can drop the first term and deal with the interaction term; because it only depends on
σ3 , the full spin up or full spin down states are ground energy eigenstates; in this case the ground state
is degenerate and corresponds to an ordered phase. We deduce that small λ corresponds to the low
temperature regime. Instead, when λ → ∞ we are in a disordered state. Indeed, in this limit we keep
the firs term, corresponding to independent spins, and any state which is a product of eigenstates of σ1
is an eigenstate. The lowest energy state is the product of the eigenvectors of σ1 with eigenvalue 1, and
therefore it is non-degenerated. In the first case (λ ≪ 1 ) the expected value (magnetization per site)
⟨σ3 ⟩ = ±1, in the second case (λ ≫ 1 ) the mean magnetization vanishes ⟨σ3 ⟩ = 0; for some
intermediate value we have a transition between the ferromagnetic and the paramagnetic states.

To diagonalize H(λ) we map the spin operators (pauli matrices) into a pair of majorana fermions c and
d , using the Jordan-Wigner transformation. A fermion operator fx , which annihilates a fermion at site x ,
can be represented by two real fermion operators (majoranas):

cx + idx cx − idx
fx = , fx† = ,
√2‾ √2‾
which satisfy the anticommutation rules,

{c(x), d(x)} = {c(x), d(y)} = 0, {c(x), c(y)} = {d(x), d(y)} = δxy .

The analogous spin operators are,

σ1 − iσ2 σ1 + iσ2
σ− = , σ+ = ,
2 2
which satisfy the mixed (boson-fermion) commutation relations,

[σ + (x), σ − (x)] = 0, {σ + (x), σ − (y)} = 0 x ≠ y

therefore the relation between the two set of operators is not trivial, even if both have the same

7 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

dimension, the two orientations of the spin, and the two values 0 or 1 of the occupation number. The
relation which ensures the correct fermion commutation rules is in fact nonlocal EX:

√2‾ [∏ ]
1
c(x) = σ1 (y) σ2 (x),
y<x

√2‾ [∏ ]
1
d(x) = σ1 (y) σ3 (x)
y<x

These formulas can easily be inverted EX:

σ1 (x) = −2ic(x)d(x)

[∏
σ2 (x) = √2‾ ( − 2ic(y)d(y)))] c(x)
y<x

[∏
σ3 (x) = √2‾ ( − 2ic(y)d(y)))] d(x) .
y<x

The product of σ3 is transformed as,

σ3 (x)σ3 (x + 1) = 2ic(x)d(x + 1) .

Combining the above expresions we obtain the representation of H(λ) in terms of the majoranaas,

∑ ∑
H(λ) = 2iλ c(x)d(x) − 2i c(x)d(x + 1) .
x x

The first term is diagonal in x but not the second; however, because of the translation invariance we can
perform a fourier transformation to k space:
π
dk ikx
∫−π
c(x) = e ck ,
2π

and an analogous formula for d(x) → dk . Replacing this expresion into H(λ) , we obtain EX,
π
dk
∫0
H(λ) = Hk (λ), Hk (λ) = 2iλ(c†k dk + ck dk† ) − (eik c†k dk + e−ik ck dk† ) .
2π

Note that the integral goes over k > 0, to explicitly exhibit the hermitian structure of the hamiltonian.
We used the formula, valid in the limit L ≫ 1:

L/2
eikx = 2πδ(k) .
∑
lim
L→∞
x=L/2+1

We can conveniently put Hk (λ) in matrix form:

)( ) ( dk )
0 2i(λ − eik ) ck
( ck ≡ C † Ĥ C ,
†
Hk (λ) = dk†
−2i(λ − e−ik ) 0

8 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

more suitable for diagonalization. The 2 × 2 matrix Ĥ has the eigenvalues ±ελ :

ελ (k) = 2√‾(λ‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾
− 1)2 + 2λ(1 − cos k)‾ (2)

we can thus define a unitary transformation U such that

( bk )
ak
Hk = C † Ĥ C = A† (U † Ĥ U)A, C = UA, A=

in the new basis A the hamiltonian becomes,

Hk (λ) = ελ (k)(a†k ak − b†k bk ) ,

whose lowest energy eigenvalue is −ε . Note moreover that ε is an even function of k . Is just this
eigenvalue that will contribute to the partition function, giving us an exact solution of the phase
transition as we see now. The explicit expression of U is EX,

√2‾ ( e−iθk 1 )
1 1 −eiθk iθk (λ) λ − eik
Uk (λ) = , e =i .
|λ − eik |

We can now compute the partition function (for L → ∞ ):

1
Z(λ) = Tr(V1 V3 )M = Tr exp[−MτH(λ)]
(cosh Kτ⋆ )L

[ ∫−π 2π k ]
π
1 dk
= ⋆ L
Tr exp −Mτ H (λ) ,
(cosh Kτ )

then, taking the trace, we obtain the explicit formula,

[ ∫−π 2π λ ]
π
dk
Z(λ) = exp −Mτ ε (k) ,

from which we compute the free energy per spin f = F/M = −(T/M) ln Z ,
π
dk ‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾
∫−π
(λ − 1)2 + 2λ(1 − cos k)‾ ,
2π √
f (λ) = f0 (Kτ ) − Jx

we took the lowest eigenvalue, recall also that τ = Kx = Jx /T , in addition, we included in the first term
the prefactor in Z , which do not contribute to the free energy near the singularity. Although the integral
is of the elliptic type, we are only interested in finding its singular behavior, which must occur when

T − Tc
t≡ =λ−1
Tc

is small (t = 0 is then the critical point). To see this we compute the specific heat near t = 1; it is given
by the second derivative

∂2 f T ∂2 f
cV = −T 2 = − 2 2 .
∂T Tc ∂t

9 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

In summary, in the neighborhood of t → 0 we may write EX,

π
dk
∫0
cV ∼ ,
√‾t 2‾‾‾‾‾‾‾‾‾‾‾‾‾
+ 2(1 − cos k)‾

(the symbol ∼ means that we keep only the dominant terms, without irrelevant factors or regular terms)
changing the variable k/t → k , we obtain,
π/t π/t
dk dk
∫0 + (2/t 2 )(1 − cos tk)‾ ∫0
cV ∼ ∼ ,
√‾1‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾‾ √‾1‾‾‾‾
+ k‾2

(cos x ≈ 1 − x 2 /2) or,

cV ∼ − ln |t| .

Therefore, the free energy is singular leading to a logarithmically diverging specific heat. This is the
signature of the ferromagnetic-paramagnetic transition at T = Tc .

The specific heat near the transition.

Exercise
Onsager found, from the computation of the spin-spin correlation at large distance, the exact formula of
the mean magnetization:
1

( sinh2 Kx sinh2 2Kτ )

1 8
⟨s⟩ = 1− ,

here written in the general anisotropic case. Show that the reparametrization

Kx = τ, Kτ⋆ = λτ

where tanh Kτ⋆ = e−2Kτ , leads in the τ → 0 limit, to the singular behavior (λ ≤ 1, in the
ferromagnetic region)
1
⟨s⟩ ∼ (1 − λ) 8 ,

10 of 11 1/21/22, 13:02
Onsager phase transition | Random physics http://verga.cpt.univ-mrs.fr/pages/PS-onsager.html

near the critical point λ = 1. Therefore, the critical exponent is β = 1/8 (different to the mean field
value 1/2).

Notes
Exercises, generally intended to fill the omitted steps of a calculation, are signaled by EX.

We followed in this lecture the book: R. Shankar, Quantum Field Theory and Condensed Matter,
Cambridge, 2017.

1. R.Voltz, V. Pierron-Bohnes, M.-C. Cadeville, Pierre Weiss (1865-1940) (.pdf (/pdfs/Weiss-

(1865-1940).pdf)) ↩

2. M. Niss, History of the Lenz-Ising Model 1920–1950: From Ferromagnetic to Cooperative

Phenomena, Archive for History of Exact Sciences 59, 267 (2005) (.pdf (/pdfs/Niss-2005.pdf)) ↩

3. Kramers, H. A., and Wannier, G. H., Statistics of the Two-Dimensional Ferromagnet, Phys. Rev.
60, 252 (1941) ↩

4. E. Fradkin and L. Susskind, Order and disorder in gauge systems and magnets, Phys. Rev. D 17,
2637 (1978) (https://journals.aps.org/prd/abstract/10.1103/PhysRevD.17.2637) ↩

5. L. Onsager, A two-dimensional model with an order-disorder transition, Phys. Rev. 65, 117
(1944) (https://journals.aps.org/pr/abstract/10.1103/PhysRev.65.117) ↩

6. T. Schultz, D. Mattis and E. Lieb, Two-dimensional ising model as a soluble problem of many
fermions, Rev. Mod. Phys. 36, 856 (1964) (https://journals.aps.org/rmp/abstract/10.1103
/RevModPhys.36.856) ↩

Archives (/archives.html) Categories (/categories.html) Tags (/tags.html)

Pelican (https://github.com/getpelican/pelican) / (https://getbootstrap.com) /

(http://creativecommons.org/licenses/by-nc/4.0/) CC 4.0

11 of 11 1/21/22, 13:02