Making Sense of The Legendre Transform
Making Sense of The Legendre Transform
Making Sense of The Legendre Transform
Related Articles
An Introduction to Statistical Mechanics and Thermodynamics.
Am. J. Phys. 81, 798 (2013)
Using computation to teach the properties of the van der Waals fluid
Am. J. Phys. 81, 776 (2013)
StefanBoltzmann law for the tungsten filament of a light bulb: Revisiting the experiment
Am. J. Phys. 81, 512 (2013)
University student and K-12 teacher reasoning about the basic tenets of kinetic-molecular theory, Part I: Volume
of an ideal gas
Am. J. Phys. 81, 303 (2013)
Predicting the future from the past: An old problem from a modern perspective
Am. J. Phys. 80, 1001 (2012)
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
Edward F. Redish
Department of Physics, University of Maryland, College Park, Maryland 20742
Susan R. McKay
Department of Physics and Astronomy, University of Maine, Orono, Maine 04469
http://aapt.org/ajp
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
614
Fig. 1. Color online The graph of a convex function Fx. The tangent line
at one point is illustrated.
is a strictly monotonic function of x because this characterization also permits us to treat functions whose negative is
convex.
A graphical way to see how the value of x and the slope of
a convex function can stand in for each other can be seen by
considering the example in Fig. 1, where the curve drawn to
represent F is convex. As we move along the curve to the
right as x increases, the slope of the tangent to the curve
continually increases. In other words, if we were to graph the
slope as a function of x, it would be a smoothly increasing
curve, such as the example in Fig. 2. If the second derivative
d2Fx / dx2 exists everywhere within the range of x in which
F is defined; part of the condition for a smooth F, there is a
unique value of the slope for each value of x, and vice versa.
The corresponding mathematical language is that there is a
615
dG
.
ds
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
615
Gs + Fx = sx.
This equation should be read carefully. Despite its appearance, there is only one independent variable: either s or x.
Referred to as a conjugate pair, these two variables are related to each other, through either xs = dGs / ds or sx
= dFx / dx. A careful writing of Eq. 8 would read either
Gs + Fxs = sxs or Gsx + Fx = sxx. To check the
consistency with Eqs. 1 and 3, we can start with, say, the
first of these equations and differentiate with respect to s. By
applying the chain rule for dF / ds = dF / dxdx / ds, we recover dG / ds = x.
slope s
sx
F
G
G
x
Fmin = G0.
B. Symmetric representation of the Legendre transform
This symmetric geometrical construction allows us to display a number of useful and elegant relations that shed light
on the workings of the Legendre transform. In particular, we
consider the symmetries associated with the inverse
Legendre transform, extreme values, and derivative relations.
Ordinarily, the inverse of a transformation is distinct from
the transform itself. For example, an inverse Laplace transform is not given by the same formula. The Legendre transform distinguishes itself in that it is its own inverse. In this
sense, it resembles geometric duality transformations.
Symbolically, we may denote this relation as:
F,x G,s.
and
Gsext + F0 = 0.
10
Gs + Fx = sx
G0 + Fxext = 0
and
dG
=x
ds
and
dF
= s.
dx
12
and
d2F ds
= .
dx2 dx
13
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
616
d 2G
ds2
d 2F
= 1.
dx2
14
d 3G d 2G
ds3 ds2
3/2
d 3F d 2F
dx3 dx2
3/2
= 0.
15
dH
dp
p
2
+ m2
16
p = mv/1 v2 .
17
18
dU
dx
19
= f.
x0
dU
,
dx
20
dV
df
21
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
617
22
23
24
25
where Pri , p i is the probability of finding the configuration of positions and momenta ri , p i. In this case, the
Hamiltonian H is explicitly given by
H = hri,p i =
i
p 2i
+ Uri ,
2m
26
E =
E Hri,p i,
27
r,p
28
EeEdE.
29
e H .
30
r,p
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
618
e H =
ehri,p i =
r,p i
r,p
drdp ehr,p
31
ZeEd ,
32
33
eF+Ed .
34
dE F
d
= 0.
0
35
= E.
0
36
37
or using Eq. 28
SE + F0 = 0E,
38
dStot
dE1
39
= 0.
E*
1
dS1
dE1
E*
1
dS2
dE2
40
,
E*
2
dS
dE
41
1E1* = 2E2*.
42
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
619
= ,
dF
d
Z = eF ,
43
= E,
44
where Z = dEeEE and F + SE = E in the thermodynamic limit. We can now see where the Legendre transform enters and why it is useful. The entropy S is a function
of E, but the internal energy is typically not easy to control.
To put more less energy into a system, we may heat cool
it. In other words, we often manipulate E by coupling the
system to an appropriate thermal bath, so the temperature or
becomes the control variable. In that case, we can perform
a Legendre transform of SE and work with F instead.
Because S , E and F , contain the same information
about the system, it makes sense to deal with the more convenient thermodynamic potential when we change the control on a system from one variable to another.
Because the independent variable in a thermodynamic potential is to be regarded as a control or a constraint parameter, the slope associated with this function for example,
dS / dE and dF / d carries physically significant information, namely, the response of the system to this control. The
Legendre transform exchanges the role of the variables associated with control and response. In the example we just
discussed, temperature or is taken as the very familiar
control variable, and the internal energy is regarded as the
response. Thus, the free energy F is the more appropriate
potential, with E = dF / d being the response. In the transformed version, which is mathematically and conceptually
easier to grasp, E is a constraint conserved variable for an
isolated system and SE is the more appropriate potential.
After we understand the significance of its slope, dS / dE, we
can identify the response as a measure for temperature.
There are many other examples of response/control pairs to
which the same kind of transformation may be applied, such
as particle number and chemical potential, polarizability and
electric field, and magnetization and magnetic field.
VI. LEGENDRE TRANSFORM WITH MANY
VARIABLES
The thermodynamic potentials depend on many variables
other than the total energy E. Each variable that can be independently controlled elicits a distinct response. As we construct Legendre transforms for each of these control/response
variable pairs, we generate a new thermodynamic potential.
The result is a plethora of thermodynamic functions. We emphasize that all these thermodynamic potentials carry the
same information but encoded in different ways. We begin
620
F
mF,
xm
45
and MM + 1 / 2 second derivatives, mF, which can be regarded as a symmetric matrix. The convexity restriction requires that all of the eigenvalues of this matrix are positive
or negative.9 In the context of thermodynamics, convexity
is the condition for stability in equilibrium systems.10 A standard corollary is that the relation between xm and sm is
one to one, so that we can replace any one of the xms by the
corresponding sm through a Legendre transform.
Because we can transform any number of the xs, we may
consider up to 2 M functions. For example, if we restrict
ourselves to E , Vthe standard variables for the microcanonical ensemble of the ideal gasthere are four thermodynamic functions: entropy, enthalpy, Gibbs, and the Helmholtz
free energy. One way to picture the relation between so many
functions is to put them at the corners of an M-dimensional
hypercube. Each axis in this space is associated with a particular variable pair xm , sm. Going from one corner to an
adjacent corner along a particular edge corresponds to carrying out the Legendre transform for that pair. For the M = 2
example of x1 , x2 = E , V, the hypercube reduces to a
square, which is related, but not identical, to the square that
appears in some texts.2,11 Thanks to the commutativity of
partial derivatives, going from any corner to any other corner
is a path independent process, so that the function associated
with each vertex is unique. For example, if we exchange
x , xm for s , sm, the Legendre transform relations would
be the simple generalization of Eq. 11:12
Fx1, . . . x, . . . xm, . . . x M + Gx1, . . . s, . . . sm, . . . x M
46
= s x + s mx m ,
47
m mHmnF = n ,
48
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
620
Let us now apply these considerations to the thermodynamics of a gas. We begin with the microcanonical partition
function E , V and consider the mapping
Fx1,x2 SE,V ln ,
49
50
51
E
and arrive at
G, + SE,V = E + V,
52
53
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
621
622
Downloaded 08 Oct 2013 to 134.129.185.248. Redistribution subject to AAPT license or copyright; see http://ajp.aapt.org/authors/copyright_permission
622