Two ways to derive the BE and FD distribution

functions
and some other guidance on statistical physics

Andrew M. Steane
Exeter College and Department of Atomic and Laser Physics, University of Oxford.

We will obtain the Bose-Einstein and Fermi-Dirac distribution functions, which are
1
⟨ni ⟩ = (1)
exp(β(ϵi − µ) ± 1

1
where the plus sign should be taken for Fermions and the minus sign for Bosons. Here ⟨ni ⟩ is the
mean occupation number of a single-particle quantum state |i⟩ whose energy is ϵi , in a perfect gas at
chemical potential µ and temperature T , with β = 1/kB T . The total number of particles in the gas is
X X
N= ni ≃ ⟨ni ⟩ (2)
i i

where the approximation is very good in practice.

In the following we first give a derivation of the above using the concept of maximisation of entropy,
under given constraints which are incorporated by Lagrange multipliers. We then give a derivation
using the grand partition function. Finally, we will point that there is also a method quite like the
second one, but which uses the canonical (as opposed to grand canonical) ensemble.

Why bother with two different methods? The Lagrange multiplier method is conceptually and alge-
braically simpler and gives the correct answer, but it uses some approximations which may not be
valid in certain cases such as very low temperature. The grand partition function method, while a lot
more involved, does not require approximations.

In all the methods, we model a gas as a collection of N particles. A quantum given state of the whole
gas is specified by stating how many particles are in each of the possible single-particle quantum states
|i⟩. Such a specification can be given by furnishing the set of occupation numbers {ni }. We then have
X
N= ni (3)
i

and the total energy of the gas is

X
U= n i ϵi (4)
i

where ϵi is the energy of the single-particle state |i⟩.

1 Microcanonical (Method by entropy and Lagrange multi-

pliers)

1.1 Fermions

Consider a gas of identical fermions. In practice there will usually be some large amount of degeneracy
gr of each energy level ϵr , and if there is not we can group states of similar energy together so that
there are gr quantum states in the r’th group.

Now suppose we want to assign nr particles to the r’th group. Since for Fermions there can be at
most 1 particle per state, the number of ways of doing this is equal to the number of ways of choose

2
nr items out of gr items (without regard to ordering), which is

gr !
Wr = (5)
nr !(gr − nr )!

Note that for fermions the ni can only be zero or one, but the nr can be large when the gr are large.

The number of ways of independently assigning particles to all the groups is

W = Πr Wr . (6)

Hence
X X
ln W = ln Wr ≃ gr ln gr − nr ln nr − (gr − nr ) ln(gr − nr ) (7)
r r

where we used Stirling’s approximation, which here is valid when

gr ≫ 1, nr ≫ 1, (gr − nr ) ≫ 1. (8)

(In practice we can usually ensure the first two conditions hold, but the third condition may not hold
at very low temperatures; we shall return to this point.)

Now the entropy is proportional to ln W and we want to find a set of nr which will maximise the
entropy subject to the constraints (3),(4), which here take the form
X X
N= nr = const, U= nr ϵr = const. (9)
r r

From (7) we have

 
∂ln W gr
= ln(gr − nr ) − ln nr = ln −1 . (10)
∂nr nr

The maximum of ln W is found when

d (ln W + αN − βU) = 0 (11)

where α and β are Lagrange multipliers (the minus sign is taken so that later on it will be found that
β = 1/kB T ). Thus we have
X  gr 
− 1 + α − βϵr )dnr = 0 (12)
r
nr

and furthermore all the nr can vary independently in this equation (the Lagrange multiplier method),
therefore each term is itself zero:
 
gr
ln − 1 + α − βϵr = 0. (13)
nr

3
 ◦ ◦ ◦ | ◦ ◦ | | ◦ | ◦ ◦ ◦ ◦

Figure 1: Partitioning a set of objects. The illustration shows a case with nr = 10, gr = 5. We
want to count ways of placing the nr objects in gr ‘pigeonholes’ when the objects are mutually
indistinguishable (and there is no restriction on the number in any pigeonhole). The diagram shows
one possible arrangement: it has {3, 2, 0, 1, 4} particles allocated to each of the 5 states. By rearranging
the nr + gr − 1 items (circles and lines) in the diagram we shall get other assignments whenever the
rearrangement does not merely permute the circles among themselves or the lines among themselves.

Solving for nr , we obtain

gr
nr = . (14)
exp(−α + βϵr ) + 1
This is the number of fermions in a group of gr quantum states, each having energy ϵr , in conditions
of thermal equilibrium, in a gas of identical fermions. The number per quantum state is therefore
nr 1
ni = = β(ϵ −µ) (15)
gr e i +1
where we introduced µ = α/β. This µ is the chemical potential, a fact that can be deduced by using
the results to deduce how much energy is required to add a particle to the system. Equation (15) is
the Fermi-Dirac distribution function (compare with equation (1)).

1.2 Bosons

We shall treat now a perfect gas of bosons. As for the previous case, we suppose the quantum states
are in groups of gr at energy ϵr , and we will assign nr particles to the r’th group. How many ways are
there to do this for bosons? To answer this question, consider a line of nr identical blobs and gr − 1
partitions separating them into subsets (Fig. 1). In this way we arrive at gr subsets, so we can see
this as a way to assign particles to quantum states: each subset in our partitioning is associated with
one quantum state. Clearly there are (nr + gr − 1)! ways to arrange the blobs and partitions. But
the associated quantum state of the whole set of particles will be the same if the blobs are rearranged
among themselves, or if the partitions are rearranged among themselves. This reduces the total count
by nr !(gr − 1)! so we find
(nr + gr − 1)!
Wr = (16)
nr !(gr − 1)!
and therefore
ln Wr ≃ (nr + gr ) ln(nr + gr ) − nr ln nr − gr ln gr (17)
where we used Stirling’s approximation and took gr ≫ 1. Hence
 
∂ln Wr gr
= ln(nr + gr ) − ln(nr ) = ln +1 . (18)
∂nr nr

4
 “microcanonical”
U, V, N fixed
U, V, N W = W (U, V, N )

S = kB ln W

“canonical”
T, V, N fixed
U fluctuates
V, N T Z = Z(T, V, N ).

F = −kB T ln Z
reservoir

“grand canonical”
T, V, µ fixed
U and N fluctuate
V T, µ Z = Z(T, V, µ)

Ω = −kB T ln Z
reservoir

Figure 2: The three main approaches to statistical thermal physics. The shading indicates isolation.
The Helmholtz function F = U − T S. The grand potential Ω = F − µN . Each approach leads to an
expression for a thermodynamic potential in terms of its natural variables, and therefore to all other
thermodynamic information.

Notice how similar this is to equation (10). The rest of the argument goes exactly as it did for fermions,
only with this sign change in the 1 term, and the result is
nr 1
ni = = β(ϵ −µ) . (19)
gr e i −1
This is the Bose-Einstein distribution function.

5
2 Grand Canonical ensemble

2.1 Probability of a microstate

In the grand canonical method we consider a system able to exchange both energy and particles with
a large reservoir, and in equilibrium with that reservoir. In this case the system and the reservoir
share the same temperature and the same chemical potential. In the basic reasoning to follow, TR
and µR refer to properties of the reservoir in the first instance, but we can employ the thermodynamic
result T = TR and µ = µR . We also introduce β = 1/kB T as usual.

The starting point of the method is to find a formula for the probability that the system will be found
in a quantum state (i.e. a single microstate) having N particles and energy E. The formula is

p(N, E) ∝ eβ(µN −E) . (20)

We will now derive this. (Note, if there is more than one microstate having the given N and E this
is ok; we will be finding the probability for the system to be found in any one microstate.)

The system and reservoir together form a composite system which is isolated, and therefore has fixed
volume V0 , energy U0 and particle number N0 . In the following, we use a subscript zero to refer to
properties of the composite, a subscript R to refer to properties of the reservoir, and no subscript
for the system. Let WR (UR , NR ) be the number of microstates of the reservoir for given macroscopic
reservoir properties UR , NR . We shall consider a single microstate of the system, one in which the
system has energy N particles and energy E. Whenever the system has N, E the reservoir must have
N0 − N and particles and energy U0 − E, so the number of microstates of the reservoir is

WR (U0 − E, N0 − N ). (21)

Since we are considering a single microstate of the system, this is also the number of microstates
available to the entire isolated composite system. On the assumption of equal a priori probabilities in
an isolated system, it follows that the probability for this situation to arise is

p(N, E) = const. WR (U0 − E, N0 − N ) (22)

SR (U0 −E, N0 −N )
= const. e . (23)

Now expand the entropy in a Taylor series:

∂SR ∂SR 1 ∂ 2 SR 2
SR = SR (U0 , N0 ) − E− N+ E + ... (24)
∂UR ∂NR 2 ∂UR2

where we have written just one of the second-order terms but we will consider all of them in a moment.
Next, from the fundamental thermodynamic relation dU = TdS − pdV + µdN one deduces
∂SR 1 ∂SR µR
= , =− . (25)
∂UR TR ∂NR TR

6
Substituting these in (24) we have
E µR N 1 ∂TR 2
SR = SR (U0 , N0 ) − + − E + ... (26)
TR TR 2TR2 ∂UR
Now we need to argue why it is legitimate to drop the second- and higher-order terms in this expression.
We observe that (∂UR /∂TR ) is the constant-volume heat capacity of the reservoir, CR . The ratio of
this second-order to the first-order term in the Taylor expansion is therefore
E
≪ 1. (27)
2TR CR
The important point is that we can always make the reservoir large enough to ensure that this quantity
is small, and furthermore it tends to zero in the thermodynamic limit for the reservoir. (To see that
the ratio is indeed small, observe that for typical states of a system in thermal equilibrium, E is at
most a few times N kB T whereas CR can be of order NR kB T , and NR ≫ N .) Similar arguments
apply to the other second-order terms which we have not explicitly written, and to the third- and
higher-order terms. Therefore in the thermodynamic limit of the reservoir, we obtain
E µR N
SR = SR (U0 , N0 ) − + . (28)
TR TR
By substituting this into (23), and using µ = µR and T = TR , equation (20) follows. Note that
although the method proceeded via a Taylor series expansion, we have argued that the result is exact
in the thermodynamic limit of the reservoir.

A remark on the reasoning via SR (U0 − E, N0 − N ).

In the above reasoning it is crucial to consider the second- and higher-order terms, not just the first-
order terms. Here is why. If we had set out from (22) we could have expanded WR in a Taylor series,
and had we kept only the first order terms in that expansion, we would have found a different formula
for p(N, E) (in fact one then gets a linear approximation to p(N, E) so it is not completely wrong,
but it is not sufficiently accurate to be of any use). The point is that the correct result requires the
whole Taylor series to be considered. We found that the linear approximation to SR is correct in the
thermodynamic limit: the entropy really does vary linearly with changes in UR and NR of the order
of E and N , whereas the number of microstates WR does not (except for unusually small values of E
and N ).

2.2 Partition function

Having obtained the crucial equation (20) we can now proceed to

1 β(µNj −Ej )
p(Nj , Ej ) = e (29)
Z
where
X
Z= eβ(µNj −Ej ) . (30)
j

7
 1
ψ(x)

−1

−0.2 0 0.2 0.4 0.6 0.8 1 1.2

x
j 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
{n1 , n2 , n3 } 400 310 220 301 130 211 040 121 202 031 112 022 103 013 004
Ej 4 7 10 12 13 15 16 18 20 21 23 26 28 31 36

Figure 3: A worked example of the relationship between states and occupation numbers. The plot
shows the first three wavefunctions for states of a single particle in a one-dimensional box (infinite
square well). Underneath are listed the 15 possible states of the whole system when there are 4
indistinguishable bosons, if we imagine they are restricted to just these three wavefunctions. The
diagram on the right is a pictorial representation of the first cases.

In these equations, j is a label for a quantum state |j⟩ in which the system has Nj particles and energy
Ej , and the sum is over all available states. The symbol Z is called the grand canonical partition
function (or just grand partition function). Note, there can be many quantum states having the same
Nj or Ej or both; (e.g. it may be that N1 = N2 = N3 and E6 = E7 or whatever); each quantum state
is included as one term in the sum.

We can use the grand partition function to find the average number of particles, for the average is
given by
X
⟨N ⟩ = Nj pj (31)
j

(where pj ≡ p(Nj , Ej )) and we have

∂ln Z 1 X X
= βNj eβ(µNj −Ej ) = β Nj pj = β⟨N ⟩ (32)
∂µ Z j j

hence
1 ∂ln Z
⟨N ⟩ = . (33)
β ∂µ
This result will be useful in finding the Fermi-Dirac and Bose-Einstein distributions. We just need to
find an expression for Z in terms of macroscopic properties of the system (that is, T, V, µ).

So far the reasoning could apply to any type of system. From now on we focus on a perfect gas. In
this context a gas is called ‘perfect’ or ‘ideal’ when it can be treated as a collection of particles which

8
 States, modes, occupation numbers
The word ‘state’ is much employed in statistical physics, and in the case of multiple particles
it can be confusing. For example, one may say there are n1 particles in the first state, n2
in the second, and so on, and at the end of such a list one has described just one state
of the entire system! For this reason I think it better to say there are ‘n1 particles in the
first single-particle state, n2 in the second,’ etc. An even better terminology is ‘n1 particles
in the first mode, n2 in the second,’ etc. Here the term ‘mode’ expresses the fact that
strictly speaking we are concerned with quantum field theory and the wave vectors k used
to describe the available single-particle ‘states’ are really describing the plane-wave modes
of a bosonic or fermionic field. The ‘occupation number’ is the state of excitation of the
mode.

do not interact, except just enough to allow them to come to thermal equilibrium. This situation is
called weakly interacting. It means we assume the quantum states and their energies are not influenced
by the movement of particles between states. In this situation there is a one-to-one correspondence
between each quantum state |j⟩ of the entire system and a set of occupation numbers {ni }. The
occupation numbers say how many particles are in the first single-particle state, how many in the
second single-particle state, and so on. Figure 3 gives an example. This is an important basic concept
so we repeat it:
|j⟩ = |n1 ⟩ |n2 ⟩ |n3 ⟩ · · ·
1-to-1
j ←→ {n1 , n2 , n3 , . . .} (34)
and
X X
Nj = ni , Ej = n i ϵi (35)
i i

where ϵi are the energies of the single-particle states. By using these expressions in (30) we find
M
!
XXX X X
Z= ··· exp β (µ − ϵi )ni (36)
n1 n2 n3 nM i=1

where M is the number of single-particle states (which may be infinite). It is important to note that
by summing over each ni we account correctly for each and every state |j⟩, without double-counting
or missing any out. Next we use the fact that an exponential of a sum is equal to a product of
exponentials:
eβµ(n1 +n2 +··· )−β(n1 ϵ1 +n2 ϵ2 +··· ) = eβ(µ−ϵ1 )n1 eβ(µ−ϵ2 )n2 · · · . (37)
Employing this in (36) we have:
XX X
Z = ··· eβ(µ−ϵ1 )n1 eβ(µ−ϵ2 )n2 · · · (38)
n1 n2 nM
" #" #
X X
β(µ−ϵ1 )n1 β(µ−ϵ2 )n2
= e e ··· . (39)
n1 n2

9
This is a remarkable result and the reader is encouraged to convince themselves that the multiple
sum does indeed factorize in this way (for example, consider the simple case (A0 B 0 + A0 B 1 + A1 B 0 +
A1 B 1 ) = (A0 + A1 )(B 0 + B 1 )).

We now have
M
Y
Z = Zi , (40)
i=1
X
Zi = eβ(µ−ϵi )ni . (41)
ni

It is tempting to ‘read’ the quantity Zi as another partition function, and so it is, as we now show.
What we need to do is realise that the overall probability of the entire system state j can itself be
written as a product:
M
Y
pj ≡ p(Nj , Ej ) = p(n1 , n2 , · · · ) = pi (ni ) (42)
i=1

where
1 β(µ−ϵi )ni
pi (ni ) = e . (43)
Zi
This factorization of the probability follows from the fact that the particles are acting independently
and furthermore the occupation numbers n1 , n2 , · · · are not constrained to add up to any particular
number. In these conditions the statistical distribution for any given single-particle state is indepen-
dent of the others. (The reader is also invited to notice that (43) is indeed the unique formula which,
upon being substituted into (42), will yield (29) for all values of µ, β.)

We have now arrived at a convenient way to proceed. We can focus on one particular single-particle
state: the one labelled by i. This state is itself a (somewhat abstract) kind of ‘system’ which can have
energy and particles. When it has ni particles it has energy ni ϵi . But how can a single-particle state
have lots of particles? The answer is one of terminology. The phrase ‘single particle state’ refers to a
wavefunction which can be that of a single particle. For a gas of many particles, the situation is that
zero or more of the particles may have that wavefunction at any given time.

There are just two steps remaining to obtain the distribution functions. First, by an argument
essentially the same as that which led to (33) we find
1 ∂ln Zi
⟨ni ⟩ = . (44)
β ∂µ
Hence if we can find an expression for ZP i then ⟨ni ⟩ will follow. But an expression for Zi is easy to
find, for (41) is a geometric series! (It is ni zini where zi = exp(β(µ − ϵi )). In the case of bosons the
series runs from ni = 0 to ni = ∞ and we find
1
Zi = bosons. (45)
1 − eβ(µ−ϵi )

10
 1.5

1
⟨n⟩

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
ϵ/µ

Figure 4: The Fermi-Dirac distribution function at 4 example temperatures. We plot the mean
occupation of a state of energy ϵ, as a function of ϵ/ϵF where ϵF is the Fermi energy (see text). Full
curve: T = 0; gray dashed: kB T = 0.1 ϵF ; red dash-dot: kB T = 0.25 ϵF ; blue dots: kB T = ϵF . In order
to plot the function at any given T , one must allow for the fact that µ falls below ϵF as T increases
from zero (see Fig. ??). One can locate µ on the graph of ⟨n(ϵ)⟩ by observing that ⟨n⟩ = 0.5 when
ϵ = µ.

In the case of fermions there are just two terms ni = 0 and ni = 1 and we find

Zi = 1 + eβ(µ−ϵi ) fermions. (46)

By substituting these expressions into (44) we obtain

1



 eβ(ϵi −µ) + 1
 fermions
⟨ni ⟩ = (47)

 1

 bosons
eβ(ϵi −µ) − 1

3 Canonical ensemble

This section can be omitted at first reading.

For systems of distinguishable particles, the canonical ensemble is simpler to use than the grand
ensemble, and this is why it is normally introduced before the latter in discussions of statistical
mechanics. However in the case of a system of indistinguishable particles the canonical approach is
difficult when we cannot ignore the indistinguishability. This is the case when occupation numbers
are not small compared to 1, and this situation arises at low temperature.

11
 2

1.5
⟨n⟩

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
ϵ/kB T0

Figure 5: The Bose-Einstein distribution function at 2 example temperatures. We plot the mean
occupation of a state of energy ϵ, as a function of ϵ/kB T0 where T0 is the BEC transition temperature
(see text). The curve for T = 0 would be a delta-function spike at ϵ = 0. Gray dashed: T = 0.5 T0 ;
red dash-dot: kB T = 2 T0 . The former case (T below T0 ) has µ very close to zero, with the result
that the ground state occupation is of order N , a large number. The latter case (T above T0 ) has
µ ≃ −0.82kB T0 with the result that none of the quantum states has a macroscopic population.

Having said that, the canonical method is not entirely forfeit. The argument given in section 2
sometimes causes students to think that in order to treat a perfect gas of identical particles we need
the grand canonical method. This is not so. If the gas has a fixed number of particles then one can
equally well use the canonical or the microcanonical approaches. Furthermore, the case of thermal
radiation can also be treated by a canonical approach since one does not need to model the radiation
as a gas of particles (with unconstrained total number of particles); one can model it instead (and
arguably more clearly) as a set of modes, each of which behaves like a quantum harmonic oscillator.

In this section we will consider a gas with a fixed number of particles, N . The canonical partition
function is
X
ZN = e−βEj (48)
j

where j labels the quantum states of the entire gas. The relationship between states and occupation
numbers (34) still holds, but now the sets of occupation numbers all add to the same total N , so (35)
is replaced by
X X
ni = N, Ej = n i ϵi . (49)
i i

We therefore have
X P
ZN = e−β i ni ϵi
(50)
{ni }

12
where the notation in the first sum is a shorthand for summing over all sets of occupation numbers
that satisfy (49).

Now the mean occupation number of state |k⟩ is

X 1 X
⟨nk ⟩ = nk pj = nk e−βEj (51)
j
ZN j

where the sum over j is the sum over all sets {ni } satisfying (49). The reader should note that this
mean value is indeed given by summing over j, not i. We use the probability pj of each state |j⟩ of
the whole gas and multiply by the value of nk for that state, then take the sum. Now observe from
(50) that

∂ZN X
= −βnk e−βEj . (52)
∂nk j

Hence we find
1 ∂ln ZN
⟨nk ⟩ = − . (53)
β ∂ϵk
This will allow us to find the distributions functions, once we have an expression for ZN .

Unfortunately the sum over {ni } in (50) is hard to do, and cannot be done exactly (in any simple
analytic way).

We now introduce a mathematical trick. Rather than trying to calculate ZN directly, we will make it
our aim to calculate the following quantity:
∞
X
X≡ e−αN ZN (54)
N =0

where α is a constant introduced for the purpose of writing this X. The idea behind the trick is that we
can calculate X (as we show in a moment) and furthermore we can estimate ZN to sufficient accuracy
by taking it that the sum determining X is dominated by one term (the largest one, obviously). That
is, we shall be taking
∗
X ≃ e−αN ZN ∗ (55)

where N ∗ is the value of N for which

∂
e−αN ZN = 0.


(56)
∂N
Using just one term (equation (55)) to estimate an infinite sum (54) can be surprising at first. In fact
this method does not estimate ZN itself very well, but it gives a very good estimate of ln ZN and all
the thermodynamic properties depend on ln ZN , not ZN directly. Section ?? shows how the step from
(54) to (55) can be justified.

13
Now we need to carefully consider the sums involved in X. We have
∞ X
X
X= (blah) (57)
N =0 {ni }

where (blah) refers to whatever we are summing (it is exp(−αN − βEj ) in this example). The N = 0
term in this double sum includes just those sets {ni } with all ni = 0. The N = 1 term includes all
those
P cases where one of the ni is 1 and the others are zero. The N = 2 term includes all those where
i n i = 2, and so on. By such reasoning one finds that the combination of sums in the definition of
X is exactly matched by another combination of sums:
∞ X
X XXX X
(blah) = ··· (blah) (58)
N =0 {ni } n1 n1 n2 nM

It follows that, if we take α = −βµ, then our X is none other than the grand partition function Z
given by (36)!

We can now use the previous discussion of the grand partition function, and thus obtain
Y ±1
X=Z= 1 ± eβ(µ−ϵi ) (59)
i

(equations (40),(45),(46)) where the plus sign is taken for fermions the minus sign for bosons. Hence
Y ±1
ZN = eαN 1 ± eβ(µ−ϵi ) (60)
i

where we used (55), taking N = N ∗ (one chooses the value of α so as to bring about this agreement
between N and N ∗ ).

Finally, by substituting our expression for ZN into the formula for ⟨nk ⟩, eqn. (53), we have
" #
1 ∂ X
⟨nk ⟩ = − αN ± ln (1 ± zi ) (61)
β ∂ϵk i

where

zi = e−(α+βϵi ) . (62)

(This zi is introduced purely in order to make the next bit of algebra easier to read.) Hence
" #
1 ∂α X 1 ∂zi
⟨nk ⟩ = − N+ . (63)
β ∂ϵk i
1 ± zi ∂ϵk

Now
 
∂zi ∂α ∂ϵi
=− +β zi , (64)
∂ϵk ∂ϵk ∂ϵk

14
therefore
( )
1 ∂α X zi X zi ∂ϵi
⟨nk ⟩ = − N− + . (65)
β ∂ϵk i
1 ± zi i
1 ± zi ∂ϵk

We will show in a moment that the term in the curly bracket is zero. Also, we have (∂ϵi /∂ϵk ) = 1
when i = k and it equals zero when i ̸= k, hence it acts to pick out one term from the sum, giving
zk 1 1
⟨nk ⟩ = = = α+βϵ (66)
1 ± zk (zk )−1 ± 1 e k ± 1

where the plus sign is to be taken for fermions, the minus sign for bosons. After substituting α = −βµ
we have the same distribution functions as before (c.f. eqn (1)).

To complete the argument, we should tidy up two points. First, the argument that the curly bracket
P
in (65) evaluates to zero is simply that the sum in that bracket can now be recognized as i ⟨ni ⟩
which must evaluate to ⟨N ⟩ and furthermore this must equal N in the canonical ensemble. The other
point is the relationship between α and µ. We argued that α is chosen so as to make the maximum
term in (54) be the one where N in the sum is equal to the number of particles actually in our system.
We then noted that as long as α = −βµ then X = Z. To check the overall consistency we now need
to confirm that this relationship between α and µ is the one which the canonical partition function
predicts. Using the expression for ZN , eqn (60), we can find µ by standard methods. We have the
fundamental thermodynamic relation

dU = T dS − pdV + µdN,
⇒ dF = −SdT − pdV + µdN, (67)

and therefore
 
∂F
µ= . (68)
∂N T,V

Using F = −kB T ln ZN one finds by an easy differentiation that µ = −kB T α, as required.

A Method to approximate an integral over a peak

In statistical thermal physics one often encounters sums of the form

X
Z= g(xi )e−αxi (69)
i

for some function g(x) and constant α. Usually the function g(x) is monotonically increasing, but not
exponentially so. In this case the terms in the sum will grow with x for small x, and decrease with x
for large x. In other words the function (g(x)e−αx ) has a peak at some x (and no other peaks).

15
 1.5

1
f (x)

0.5

0
0 5 10 15
x

Figure 6: The full line shows some peaked function f (x). We would like to estimate the area under
the curve. It can be done by fitting a gaussian curve (shown dashed) to the peak, and calculating the
area under that.

A good estimate of the sum can often be obtained by approximating it by an integral. With this in
mind, let’s consider the general task of evaluating any integral of the form
Z ∞
z= f (x) dx (70)
0

where the function f (x) has a single peak at some positive x. The idea is to fit a gaussian curve to
the peak, and then use the integral under the gaussian curve as our approximation to z (see Fig. 6).

If the fitted gaussian curve has a height f (x0 ) and standard deviation ∆x then the result is
√
z ≃ f (x0 )∆x 2π. (71)
The value of x0 is obtained by finding the location of the peak: it is given by the value of x at which
df/dx = 0. A suitable value for ∆x is obtained as follows.

Let y = ln f . Then
dy d2 y
= f ′ /f, = −(f ′ /f )2 + f ′′ /f. (72)
dx dx2
It follows that the Taylor expansion of y around the peak at x = x0 , up to second order, is
1 f ′′
y ≃ y(x0 ) + (x − x0 )2 . (73)
2 f
Hence we find, for values of x near the peak,
f (x) ≃ f (x0 ) exp(−(x − x0 )/2∆x2 ) (74)
where
p p
∆x = −1/y ′′ = −f /f ′′ (75)
(and this is to be evaluated at x = x0 ). This furnishes the information we need to evaluate our
approximation to the integral, (71).

16
A.1 Stirling approximation

As an example of the method, let’s consider the factorial function:

Z ∞
n! = xn e−x dx. (76)
0

This equation is exact. But let’s get an approximate evaluation of the integral using the above method.
We have f (x) = xn e−x and therefore y(x)√ = n ln x − x. Then dy/dx = 0 at x = n and at this value
of x, y ′′ = −1/n. Hence we find ∆x = n. The method yields
√ √
n! ≃ f (n)∆x 2π = nn e−n 2πn. (77)

This is called Stirling’s approximation. It is a fairly good approximation to n!, and it yields a very
good approximation to ln(n!):

ln(n!) ≃ n ln n − n + (1/2) ln(2πn). (78)

A.2 The precision of approximation of logarithms

For n ≫ 1 the final term in Stirling’s approximation for ln(n!) (78) is entirely negligible compared to
the others, so then we have

ln(n!) ≃ n ln n − n (79)

which is how the approximation is often used in practice. And this illustrates a rather remarkable
fact. The latter result asserts that n! ≃ nn e−n , in other words
Z
xn e−x dx ≃ (xn e−x evaluated at x = n). (80)

In this expression, the integral over all of x is being estimated by the maximum value of the integrand.
But one should be cautious of such an estimate. In fact by dropping
√ the ln(2πn) term in (78) we have
that our expression for ln(n!) is now inaccurate
√ by about ln( 2πn) and therefore the expression on
the right of (80) is inaccurate by a factor 2πn, which is a large factor for n ≫ 1.

What is going on? The above is an example of a general fact about very large numbers and logarithms.
To make the point, a numerical example suffices. Suppose we have some quantity W whose true value
is W = exp(1020 ) (note, this is an extremely large number!). Suppose now that we have a method to
estimate W , but unfortunately our estimate is wrong by a factor one million. Ordinarily one would
think that an estimate wrong by a factor one million is entirely useless for scientific purposes. But it
is not so! For, what if the only way W influences the physical effects we wish to calculate is through
its logarithm. Then we have that the true value is

ln W = 1020 (81)

17
and our faulty estimate gives

ln(106 W ) = 6 ln 10 + 1020 .

In this example, therefore, the imprecision in our estimate of ln W is just 6 ln 10 ≃ 13.8 for a quantity
whose magnitude is 1020 . This is an accuracy of about one part in 1019 , so very precise indeed.

We are now ready to justify the step from (54) to (55), in which a sum was approximated by a single
term in the sum. One first argues that the sum can be estimated
√ by an integral, and then one applies
the reasoning leading to (71). Finally, one drops the ∆x 2π factor, which is valid in the case where
one is really interested in the logarithm of the sum in the end, as opposed to the sum itself. In the
case of (54), ∆x corresponds to ∆N , the size of the thermal fluctuations in the particle number, and
the result
√ is accurate when this is small compared to N . In a wide range of circumstances one finds
∆N ≃ N which is indeed small compared to N when N ≫ 1.

18