1

Unit 2-9: Equivalence of the Microcanonical and the Canonical Ensembles in the Thermodynamic
Limit

When we introduced the Helmholtz free energy A(T, V, N ) in Notes 1-5, it was defined mathematically as the Legendre
transform of E(S, V, N ), or equivalently, −A/T was the transform of S(E, V, N ). In Notes 2-4 we saw that, in ensemble
theory, S(E, V, N ) is determined from the number of states Ω(E, V, N ) in the microcanonical ensemble at fixed energy
E. We then called Ω(E, V, N ) the microcanonical partition function. We can therefore denote the Helmholtz free
energy that we get by taking the Legendre transform of S(E, V, N ) = kB ln Ω(E, V, N ) as the microcanonical Helmholtz
free energy, Amicro (T, V, N ).

In our discussion of the canonical ensemble in Notes 2-8, we obtained the Helmholtz free energy from the canonical
partition function, A(T, V, N ) = −kb T ln QN (T, V ). We can denote this as the canonical Helmholtz free energy. How
do we know that the microcanonical Helmholtz free energy Amicro and the canonical Helmholtz free energty A are the
same?

In Notes 1-6 we motivated the physical meaning of the Helmholtz free energy by reference to a system in contact
with a thermal reservoir, in a discussion similar to that which motivated the canonical ensemble. Here we will show
explicitly that Amicro = A in the thermodynamic limit N → ∞.

Energy Fluctuations

Before we show that Amicro = A, we first consider the behavior of the energy in the canonical ensemble. In the
microcanonical ensemble the system energy E is fixed. In the canonical ensemble the system energy fluctuates, with
an average hEi that is fixed by the temperature T . In Notes 2-4 we argued, for the case of an ideal gas divided in half
by a thermally conducting wall, that such fluctuations become negligible as the size of the system gets large. Here we
will explicitly demonstrate this for a system in the canonical ensemble.

Consider the variance of the energy in the canonical ensemble. Using the canonical probability density for the system
Ω(E) −βE
to have energy E, P(E) = e , we have,
∆E QN
 2  Z Z 2
Var[E] ≡ E − hEi = hE 2 i − hEi2 = dE P(E) E 2 − dE P(E) E (2.9.1)

Z Z 2
dE Ω(E) −βE 2 dE Ω(E) −βE
= e E − e E (2.9.2)
∆E QN ∆E QN

Consider now the derivative,

Z  Z Z 
∂hEi ∂ dE Ω(E) −βE dE Ω(E) −βE dE Ω(E) −βE 1 ∂QN
= e E = e (−E 2 ) − e E (2.9.3)
∂β ∂β ∆E QN ∆E QN ∆E QN QN ∂β

where the second term follows because QN (T, V ) depends on T and hence on β = 1/kB T , and so ∂(1/QN )/∂β =
−(1/Q2N )(∂QN /∂β).

Now from Eq. (2.8.19) of Notes 2-9, we have,

1 ∂QN ∂
= ln QN = −hEi (2.9.4)
QN ∂β ∂β

So Eq. (2.9.3) then becomes,

∂hEi
= −hE 2 i + hEi2 = −Var[E] (2.9.5)
∂β
 2

Therefore,
∂hEi ∂hEi ∂hEi
Var[E] = hE 2 i − hEi2 = − =− = kB T 2 = kB T 2 CV (2.9.6)
∂β ∂(1/kB T ) ∂T
where CV = (∂E/∂T )V,N is the specific heat at constant volume. This gives a sometime useful formula for the specific
heat at constant volume in the canonical ensemble,

1  2
hE i − hEi2

CV = 2
(2.9.7)
kB T

The main point is that Var[E] scales as an extensive variable, since E is extensive and β = 1/kB T is intensive. So
Var[E] ∼ N , where N is the number of particles and is a measure of the size of the system.

Now the width of the probability density P(E) is given by its standard deviation σE ,
p p √
σE ≡ Var(E) = hE 2 i − hEi2 ∼ N (2.9.8)

and the energy is extensive so hEi ∼ N . So the relative width of the probability density, which is a measure of the
significance of the energy fluctuations, scales with the system size as,
√
σE N 1
∼ ∼ √ →0 as N → ∞. (2.9.9)
hEi N N
So in the thermodynamic limit N → ∞, the relative fluctuations of the energy in the canonical ensemble vanish.
This is one indication that the canonical and the microcanonical ensembles should give equivalent results in the
thermodynamic limit.

Equivalence of A and Amicro

We now investigate the effect that energy fluctuations have on the canonical Helmholtz free energy A, as compared
to the microcanonical Helmholtz free energy Amicro .

Microcanonical Amicro

To compute Amicro we:

1) Compute S(E, V, N ) = kB ln Ω(E, V, N ) from the microcanonical partition function Ω(E, V, N ).
2) Take the Legendre transform of S with respect to E to get (−Amicro /T ) = S − (E/T )

We can also write the Legendre transform as follows:

 
−Amicro (T, V, N ) E
= max S(E, V, N ) − ⇒ Amicro (T, V, N ) = min [E − T S(E, V, N )] (2.9.10)
T E T E

If Ē is this minimizing value of E, then we have,

Amicro (T, V, N ) = Ē − T S(Ē, V, N ) (2.9.11)

Canonical A

To compute A we first compute the canonical partition function QN (T, V ) and then take A(T, V, N ) = −kB T ln QN (T, V ).

Consider now the computation of QN = e−A/kB T .

 3

Z
dE
QN = e−A/kB T = Ω(E, V, N ) e−E/kB T use S = kB ln Ω (2.9.12)
∆E

Z
dE S(E,V,N )/kB −E/kB T
= e e (2.9.13)
∆E

Z
dE −[E−T S(E,V,N )]/kB T
= e (2.9.14)
∆E
Consider the exponent E − T S(E, V, N ) and expand to second order about its minimum Ē. With E = Ē + δE, and
recalling that (∂S/∂E)E=Ē = 1/T since Ē minimizes E − T S, we have,
   2 
∂S 1 ∂ S
E − T S(E, V, N ) = Ē − T S(Ē, V, N ) + δE − T δE − T δE 2 (2.9.15)
∂E E=Ē 2 ∂E 2 E=Ē

   
1 1 ∂(1/T )
= Amicro + δE − T δE − T δE 2 (2.9.16)
T 2 ∂E E=Ē

 
1 ∂T 1 1
= Amicro + δE 2 = Amicro + δE 2 (2.9.17)
2T ∂E V,N 2T (∂E/∂T )V,N

δE 2
= Amicro + (2.9.18)
2T CV
Using Eq. (2.9.18) in (2.9.14), we can now compute,
Z
−A/kB T dδE −Amicro /kB T −δE 2 /2kB T 2 CV
QN (T, V ) = e = e e (2.9.19)
∆E
Z
dδE −δE 2 /2kB T 2 CV
= e−Amicro /kB T e (2.9.20)
∆E

The integrand in the integral over δE has the form of a Gaussian

p which
"

E
is
√ sharply peaked at δE = 0 (i.e. E = Ē) with a width hδE 2 i =
next
µ
width
2
kB T CV . Since CV ∼ N and Ē ∼ N are both extensive, the relative n
TN
width of this Gaussian is,
p √
hδE 2 i N 1
Ē
∼
N
∼ √ →0
N
as N → ∞. (2.9.21)
p
,
Since the contribution to the Gaussian integral comes almost entirely ENN
from the range of δE within a few widths about the peak, and the width
is a very small fraction of Ē, it is therefore an excellent approximation to take the limits of integration on δE to be
±∞ (rather than −Ē and +∞) and explicitly do the Gaussian integral.
Z ∞
2 √
/2σ 2
Recalling the normalization of a Gaussian, dx e−x = 2πσ 2 , we get,
−∞
√
−A/kB T −Amicro /kB T 2πkB T 2 CV
QN (T, V ) = e =e (2.9.22)
∆E
Taking the natural log of both sides then gives,
√
2πkB T 2 CV 2πkB T 2 CV
  
kB T
A = Amicro − kB T ln = Amicro − ln (2.9.23)
∆E 2 (∆E)2
 4

So

2πkB T 2 CV
 
kB T
A − Amicro = − ln (2.9.24)
2 (∆E)2

Now since A and Amicro are both extensive ∼ N , and CV ∼ N is also extensive, and ∆E ∼ E/N is intensive, then
we have the relative difference in the canonical and the microcanonical Helmholtz free energies,
A − Amicro ln N
∼ →0 as N → ∞. (2.9.25)
A N
We conclude that A = Amicro , and hence the canonical and the microcanoncial ensembles will give the same thermo-
dynamic results, provided we are in the thermodynamic limit N → ∞.

Because of this equivalence, although we developed the canonical ensemble as a description of a system with fluctuating
energy in contact with a thermal reservoir, we could also just as well use it to describe a system at fixed energy E in
thermal isolation from its surroundings. One just has to choose the temperature T of the canonical ensemble so that
it gives hEi as the fixed energy E. If we are in the thermodynamic limit, the fluctuations of E away from hEi will be
negligible and will not effect the thermodynamic properties.

Note, the approximation we made to evaluate the integral in Eq. (2.9.14), where we expanded the argument of the
exponential to second order about its minimum and then did the resulting Gaussian integration, is known as the saddle
point approximation. We will soon see that Stirling’s formula for ln N ! is a result of a saddle point approximation.