15

Ann. d. Physik, 3, 1055-1101 , 1929 15

4. On the Kinetic Theory of Thermal Conduction in Crystals

(With 1 figure)

Introduction

Since the work of Born and v. KArmsn and others, all problems of the kinetic
theory of solids have been solved at least in principle, so long as they are concerned with
the equilibrium state. On the other hand, the treatment of non-stationary and quasi-
stationary problems still presents difficulties. The reason for this lies in the fact that for
their solution one needs to consider those processes (in general of a secondary nature)
which produce a statistical equilibrium.
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

The circumstances are exactly the same as in the kinetic theory of gases: the
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

phenomena that originate from inter-molecular collisions are of secondary importance for
setting up the equation of state. By contrast, they play a decisive role in the case of
thermal conductivity and for all those phenomena that are connected with the concept of
a mean free path.
The situation is the same for solids. Since the work of Bornl and Debye2, it has
been known that the idealised model, with harmonic forces between atoms, is useless for
an investigation of thermal conductivity, since it would lead to an infinitely large
conductivity. For in the model with harmonic forces, as we shall see in section 1, one can
think of the motions of the atoms as built up from mutually independent sound waves.
Once an arbitrary energy distribution of these vibrations exists, it will exist forever. A
thermal equilibrium will then not be established, and in general one can therefore not
speak of a temperature at all. But even a definition of the thermal conduction in terms
of the energy gradient instead of the temperature gradient would still not lead anywhere,
since the concept of thermal conductivity implies a proportionality relation between
energy flux and energy gradient, which is absent in this model. This is most easily seen
by considering states which possess an energy flux but no gradient, e.g. when there is only
a single wave present. Following Debye, one can say that the thermal conductivity of
such a model is infinitely large.
These facts have been questioned by a number of people. In particular,
Schrodinger3 tried to prove that the harmonic model produces a finite conductivity. His
observations were disproved by Ornstein and Zernike4, who demonstrated the correctness
of Debye's and Born's argument.
Recently, v. Karm66n5 tried to rescue the harmonic model in a paper which used
quantum mechanics. But his results are based on a false conclusion. v. KArman assumes
that quantum-mechanically an energy exchange between two coupled oscillators occurs
only when the excitation energies of the two oscillators differ by a single quantum (the
oscillators can of course be looked upon as representing two neighbouring atoms). In
reality, the quantum-mechanical equations always allow for the possibility that a quantum
can be passed from the one oscillator to the other, even for arbitrary differences between
the two quantum numbers. Quantisation therefore changes nothing in the final result for
this model.

I M. Born , Phys. Z., 15, 191, 1941

2 P. Debye, Vortrs,ge tuber die kinetische Theorie..., p. 43 (publ. Teubner, 1914).
3 E. Schr& finger, Ann. d. Physik , 44, 916, 1914.
4 L.S. Ornstein & F. Zernike, Proc. Amst. Acad. 19, 1295, 1925
5 T. v. Karm4n, Naturw. 17, 385, 1929
 16

16 On the Kinetic Theory of Thermal Conduction in Crystals

Moreover, Debye has already observed that one must go one step further in the
expansion of the energy as a function of the distance between two atoms, and take into
account the cubic terms, which produce quadratic terms in the equations of motion. One
can still analyse such a model in terms of plane harmonic sound waves , but in such a case
the amplitudes of the individual waves are no longer constants, but change slowly with
time. This will then give rise to a statistical equilibrium between individual waves. Each
energy distribution which is associated with a non -zero energy flux will in general no
longer be stationary. An energy transport can then only be maintained in the presence
of a temperature gradient . One should therefore expect a finite thermal conductivity for
such an anharmonic model. Debye also pointed to an essential property for this model:
the perturbations acting on a wave depend on the existence of a second wave interacting
with the first , since the equations of motion are of second order. But since at high
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

temperatures each of the waves is excited with an intensity kT, the time taken by one
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

wave to get into equilibrium with the others, will be proportional to 1 IT . The same holds
for the distance covered by a wave packet in such a time , the so-called mean free path of
the waves . It is then easily seen that for high temperatures the thermal conductivity
must be proportional to 1/T, in agreement with experiment.
Certain parts of Debye's paper , however, contain assumptions which have to be
ruled out, as was shown by Ornstein and Zernike (l.c.). For Debye considered the
scattering of waves by static deformations of the lattice. After all , the effect of the
anharmonic terms is determined by the dependence of the elastic constants on the distance
of a particular location in the lattice from its rest position . In analogy to the familiar
calculations on density fluctuations , Debye looked schematically at the effect of the
totality of waves on the one wave under consideration . He computed the mean
deformation at each point and then considered the scattering of the waves at such
deformed positions in the crystal . But the crucial point here is precisely that the waves
causing these deformations , and thus also the deformations themselves , should propagate
with the very same velocity as the wave which is supposed to be scattered by them; the
Debye approximation is therefore certainly not admissible.
Ornstein and Zernike then tried to perform the calculation correctly , i.e. they tried
to see how the energy is transmitted from one vibration to another as a result of the
additional quadratic terms in the equations of motion . For this, they used as a model a
one-dimensional elastic continuum . It then turned out that the above assumptions are not
sufficient ; the energy flux remains unchanged when two waves interact , so that there will
always exist stationary states with non -zero energy flux, and therefore the thermal
conductivity is not finite.
In what follows it will be seen that this result is not only valid for the one-
dimensional case, but also for the three-dimensional continuum ; what is more, this holds
not only to a first approximation , but rigorously and to arbitrarily high powers of the
potential function . The continuum is therefore quite unsuitable , since its behaviour is
qualitatively quite different from that of an object with an atomic structure.
The radically different behaviour of the atomic lattice was already demonstrated
by the calculations of Paulhl . However , he used a linear chain which , again , did not
produce a finite thermal conductivity . The reasons for this will be discussed in more detail
in sect . 6. For a linear, discontinuous , chain this result certainly does not hold quite so
generally : already in the next approximation the conductivity becomes finite, although
with a wrong temperature dependence.

1 W. Pauli, Verh. d. Dtsch. Phys. Ges. (3) 6, 10, 1925

 17

Ann. d. Physik, 3, 1055- 1101 , 1929 17

Thus it will be first of all our task to find an explanation for this peculiar behaviour
of the continuum. This will then be used to determine which of the properties of the
lattice (i.e. the three-dimensional lattice, as one is led to suspect from Pauli's calculations)
will have to be taken into account in order to produce a finite thermal conductivity.
It was previously mentioned that the problem of thermal conduction is very closely
connected with the establishment of the statistical equilibrium between different
vibrations; in sect. 1 we shall therefore discuss under what circumstances such an
equilibrium will be produced for certain, i.e. when it is that an H-Theorem will hold for
such lattice vibrations.
In this we shall restrict ourselves completely to processes in non-conducting, ideal,
crystals. No assumptions are thus made about the validity of processes in crystalline or
amorphous substances, which are much easier to deal with, or in metals, for which the
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

presence of conducting electrons demands special considerations'.

Before setting up the relevant formulae and notation, we shall first consider the
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

situation in a crystal with purely harmonic forces.

1. Harmonic forces2

The notation used here is illustrated by the linear model; generalising to several
dimensions then becomes an easy matter. Let us consider a chain of N atoms whose
spacing in the equilibrium state is a. Let the displacement of the nth atom from this
position be un. For boundary condition we shall choose the periodic condition, commonly
used in lattice theory:
un+N = un
The energy as a function of these displacements is of the form:

(1) E=m y[u 2+0)02( un -U._,)2] .

2

This leads to the equations of motion,

2
(2) ii" = WO (un+l + un-1 - 2un) .
At a later stage we shall need an expression for the energy flux. For this we compute the
energy in a certain interval n', n",

[
2 m ft' un2 un - U,1)2
(3) E, = + 22 ( + 22
( un - un+1)2^

Differentiating with respect to t:

n" 2
my[unun + (un - U,1 0. - un-1)
(4) n
2
0)2
+ 2 (un+1 - un ) ( un+l - u.)]

1 Note in proof: Calculations for this case will shortly be published in this journal.
2 The following considerations are in the main taken from a paper by M. Born and T. v. KArm£n,
Phys. Z. 18, 297, 1912.
 18

18 On the Kinetic Theory of Thermal Conduction in Crystals

Because of (2),

Wo2[( u n"+ 1 +un")(U. "+1-un" )- ( it +26n -1)(un - un -1)1 •

2

Thus one must evidently define

2 un + un-1
(5) S=mWO 2 (un-un-1)

as the energy flux.

To solve (2), we introduce the following convenient variables (normal coordinates)
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

(6) U. = 4f efn'
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

Here, f takes on the values

(7) f=N(0,±1,± 2,...),

because of the boundary condition. For un to be real, we must have

(8) 4-f = 4f

Formulae (1) and (5) are not the most general. One could also take into account
forces between atoms at distances a, 2a, etc. But it can easily be shown that such forces
do not introduce anything that is essentially new, since the form of the trial solution (6)
remains unaltered as long as all atoms are considered on an equal footing. To avoid
unnecessary complications, we shall therefore keep using formulae (1) and (5) in what
follows.
It is important to stress that f in (6) is only defined to within mod 2,c. This is a
result of the atomic structure of our model. (6) represents a type of wave for which the
wavelength does not have any particular importance, whereas the phase difference
between two neighbouring atoms does. A wavelength which is less than a has obviously
no meaning. If the crystal is then replaced by a continuum, (6) will again give a result,
but one would then have to consider n as continuously variable, and correspondingly let
f take on all values between -- and +-.
Apart from that, the greater N, the more densely the values off will be distributed.
In general we shall consider f as a continuous variable.
From (1) and (6) one obtains, using a simple transformation,

(9) E = ^[1c_j + wf2 f -fj,

where M = Nm, and where

(10) wf=2oblsin2

Since the transformation (6) does not contain the time, the momentum conjugate to 4f
is P4f _ = M4-f
f
 19

Ann. d. Physik, 3, 1055-1101, 1929 19

Thus, the equations of motion become

1 1 dE 2
f h -f Ma _f =-Of f

and their solutions are

4f = b1,1e'0)'t + b f._1e-'0,1

with b_ f _ j = b f, j , because of (8). The index j distinguishes between waves running to the

right and to the left. To separate these, one should observe that
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

bf,1= I(4f + 1 tf)e -'wf1'

f
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

bf,-1 = 2(4f - - tf)eQ'I1.

i1of

One can interpret (11) as a contact transformation. The momentum associated with bf1
will then be

pbf j = -iMwf jb-f. j ,

where, for formal reasons, we define wf,_1 =-wf,1 . Then

(12) 1 I
pgfd f -H= p,,..dbf,j,
where H arises by expressing the 4 f in terms of the pp .
In the variables (11), therefore, the Hamiltonian is identically zero according to
(12), so that, trivially,

bf,j = pb,; = 0 .
Finally, we also need the energy and energy flux as functions of the b. It
immediately follows from (9) that

( 13 ) 2

Furthermore, the spatial (or temporal) mean of the energy flux is, from (5),

37 = 2 1 wo2w f4 cos 2 b1,, b_

From (10), this can then be written

s = N , Efj ad
f

2
Ef,j = Mbf,j b-f,-j wf,j .

Eq.(14) holds independently from the special dispersion law (10), sincef is the group
velocity.
 20

20 On the Kinetic Theory of Thermal Conduction in Crystals

In three dimensions we want to restrict ourselves to the case of a cubic lattice and
assume that the crystal is in the shape of a cube with side Na. Let the individual atoms
be labelled by n1, n2, n3i and x'hns'ns , yn,n$ne , z, ,, ,, be their displacements from the rest
position. In place of (6) we obtain

+gn,+hny)
xn,r, na - Jghei(

(6) = n JBhei( fn,+gn'+hn^)

yn,nsns
= ^j ^f9he'(jn,+9n,+hn,)
zn,nsna

The potential energy is a quadratic form in x, y, z , analogous to (1). Its exact form is for
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

the present not of interest'. Because of (6'), it separates into a sum of quadratic forms,
each of which only contains the three 4, n, ^ for a particular index triplet fgh. Thus, by
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

means of an orthogonal transformation,

4fghj = af9hj4f9h + Ifghj 1fgh +YfgkJ fgh (j = 1,2,3),

one simultaneously diagonalises all these quadratic forms, where the coefficients a, 0, y
are different for each f, g, h . Finally, the energy is obtained in the form

(9) E_ 2
,I [tf9hj, - f 9,_ h +Ct)
f9h9
z
r9hj -r, 9,-h•7]'
fgh j=1

where now M = mN 3 and the w fghj are found to be the eigenvalues of the quadratic form
2E/M. The (fghj will in general still depend on j, and in the limiting case of a continuum,
j distinguishes the transversal from the longitudinal waves.
Again we set

bfghj = ( fghj + 1 fghj )e-^Idi

lWfghj

with w fgh, j = -w191 j and obtain

(12) bfghj=0;

(13) E _ j Efghj =
Mj (fghj b19 hj
b-f,-9,-h. -j ;
and for the mean energy flux,

awfghj
S• ILEf9 d f
N

(14') Sy _ 1
E, °?(fghj
N fghj dg

Sz Ef9h awfghj
7 ah

1 Cf. e.g. the formulae given in Born and KRrm&n, i.c.

 21

Ann. d. Physik, 3, 1055- 1101, 1929 21

2. Perturbation energy

As can be seen from the formulae of the previous section, the bf9hr are independent
of each other; if therefore there exists a non-zero energy flux at some time, then it will
continue for all times, even in the absence of an energy gradient. And this corresponds to
an infinitely large thermal conductivity. We now have to take into account that (1) is not
the complete expression for the energy. It is just the starting term of a power series. For
the case of a linear chain, e.g., its next term will have the form

(16) V = E W02 (Un - u n_1)3.

In general it will be an expression of third degree in the displacements, in which all lattice
points are treated equally, and which can be written
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

V= Fkl (n' - n, n" - n) xiaxkri xln" ,

n,n ,n",i,k,l

where n stands for n1, n2, n3 and x1, x2, x3 for x, y, z . If one now replaces the x, y, z by
the b, in accordance with (6'), (6") and (11), one obtains a factor

in the term which is multiplied by bf91,jbrVh bf"fh . One can then carry out the
summation over nl n2 n3 . Then the sum is only not equal to zero when

f+ f,+ f" =
f ±2x
(17) 9 +9 +9 0
1 ±2x
h+h'+h"{02s

It is of importance that this spatial interference condition is only valid to within mod 2ic,
since the n are integers. For the case of the continuum, the right-hand sides of the
equations (17) are only equal to 0. The additional term in the expression for the energy
therefore has the form

,f+f.+r)o
mo
(18) V = 31 Vff .b Ib 1• b1" ei

Here, and in what follows, we have written f for f g h j whenever there is no danger of
ambiguity ; Vfjf" vanishes whenever the indices do not satisfy the condition (17).
Furthermore , Vf;f" is symmetric in the three indices . - From this property it follows
immediately that the continuum model cannot be used. To see this , consider the quantity

(19)
J= I
fehi
f
f wf9hJbf9hib- -B-h-A

and its time derivative.

 22

22 On the Kinetic Theory of Thermal Conduction in Crystals

The effect of the perturbation energy is that the b remain no longer constant, but
change according to
-i _ dy
bffh,9 = pb_l -a : i Mw - db_ _ h .
f9h,7 f 9- .7
(20) bf"et(w^.+w^.-w^)t
= Mcuf
2 ^ V_f,f',f"bf,

Thus, from (19),

Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

ei(wf.+wf.-wf)e
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

J = fwf M^f V ff',f"bfbrbf"

f,f,f"
MW f Vf, f',-f"bbb f'b f"et((O-m -Wf-)t
2i es(wf +wr+wf-)t
- - M jf(Vf,f',f" bf b1bI")

Because of the symmetry of Vff f . this can also be written

- 2a bf"ei(w,+wf.+wf.)t
(21) ,(f+f'+f^^)Vfff„bfbf,
3M

This expression is only non-zero when the lower equations in (17) are satisfied. In
particular, (21) always vanishes for continuum calculations. For the one-dimensional case
this leads directly to the thermal conductivity being infinitely large, since in that case wf
= const.f and J then becomes proportional to the energy flux. This therefore represents
no change from the harmonic model. In the three-dimensional continuum, J is
independent of the energy flux, because of the dependence of the frequency on j. But the
fact that an integral of the equations of motion exists leads to states that are stationary
under the influence of the perturbation (18) and that are different from the
thermodynamic equilibrium state. They are obtained by maximizing the entropy while
keeping J fixed. Since J is asymmetrical (its sign changes when j and -j are
interchanged), a non-vanishing energy flux will then also be associated with non-vanishing
J for such states1.
Roughly speaking, the integral J is analogous to the total momentum of a gas since
the latter, too, represents an integral of the equations of motion. Our case therefore
corresponds to a gas that is contained in a circular closed pipe. In that case, too, one can
no longer speak of thermal conductivity for the gas flow through the pipe. The only
difference in the case of a gas is that one can choose the boundary conditions in such a
way that no mass transport takes place at the boundary surfaces, so that the total
momentum necessarily becomes zero. For a crystal, on the other hand, the boundary
conditions permit the passage of an energy flux through the boundary surfaces (or else

1 For this reason it would appear obvious that Ornstein and Zernike's calculations for the linear model
led to a negative result, whereas the proof for the three-dimensional case was by no means straightforward.
 23

Ann. d. Physik, 3, 1055- 1101 , 1929 23

thermal conduction would not be present); thus J = 0 is always consistent with the
boundary conditions. It is therefore important to note that J, even though it is not
identical with the energy flux, is nevertheless constructed very similar to it, in contrast
to the case for gas, where the momentum is linear but the energy flux is of third order in
the velocity.

3. Statistics for the classical calculation.

The foregoing leads to the conclusion that one has to use the discontinuous model
as a basis for further calculations. This suggests that one starts in analogy to the normal
procedure in the kinetic theory of gases : one can consider a certain wave f, g, h, j and
assume that all the others are in thermal equilibrium. One then considers the perturbation
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

energy (18). As will be seen in more detail in the next section, this will cause the wave
under consideration to be attenuated at the expense of the others. Equally, the energy
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

flux generated by this wave will decay, though perhaps more slowly since the waves
produced in the scattering processes could carry along an appreciable amount of energy
flux in the same direction as well . In this way one obtains the decay during a short time
interval, since the change produced by (18) will have to be treated as small . By analogy
to the method of the kinetic theory of gases , one would now assume that this decay
follows an exponential law. One could thus define a half-life and, by multiplication with
the group velocity, a mean free path for the waves. This then leads immediately to the
thermal conductivity, using e.g. the method of Debye (l.c.).
But such a procedure is certainly not permissible in our case. For it would produce
a finite thermal conductivity even in the case of a continuum, since the energy flux
derived from a single wave will start to decay when the perturbation energy is switched
on. But we know from the results of the previous section that it will not decay to zero.
Instead, the final state will be associated with a non-vanishing energy flux. Assuming
exponential decay is therefore not justified here.
Instead, we shall have to calculate exactly the change in the statistical distribution
due to (18). We shall choose as variables the quantities Effhj introduced by (13') and we
shall consider all values of the phases of b f9hj as equally probable. By this we shall restrict
ourselves to phenomena that are independent of position, which is sufficient for the proof
of the H-Theorem. Special ways will later have to be used to deal with thermal
conductivity.
Let us then consider the totality of similar systems and that fraction
W(Ef9hj ) dE ...
of the systems which contains the excitation potential of all the vibrations in the interval
(Ef, Ef+dEf). During time r, every system will then pass into a state Ef + 4Ef . Now let
F(E, 4E, r)d4E be that fraction of all systems that were in state E initially and for which
the 4Ef lie in the interval under consideration. The change in distribution will evidently
becomes

W(Ef,r)-W(Ef,O)= f ... f W(Ef-AEf,0)

(22)
F(Ef -1Ef,AEf,,r)dAE
f•••f W(Ef,0)F(Ef,dEf,a)dAE.

What follows is a slight modification of the well-known Fokker Method. S.A. Fokker, Ann.d.Phys.
43, 812 , 1914 and M. Planck, Berl. Ber. 1917, 324.
 24

24 On the Kinetic Theory of Thermal Conduction in Crystals

It should be noted that the E vary continuously, so that for the T sufficiently small, F will
only be appreciably different from zero when the AE are small. One can then put

W(E - 4E)F(E - LE, 4E, T)

= W(E)F(E, 4E,T) - YtEf /-[W(E)F( E LE,T)]

(23)
f f

[W(E)F(E, 4E,T)] .
+2
1AEfAEf
aEf dEf
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

(22) then becomes

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

(24) 4W=-^aEf[AE1W(E)]+2aE faEf{AEfdEf.W(E)]

with

AEf = f ... f LEJF(E,LE,T)dAE

(25) and
aTf _ = f ••• f Li fzE, f.F(E, tE, T)dztE.
AEf,

In (23), only the first two terms were written down in full detail; it will be shown later
that higher-order terms make no further contributions. The quantities (25) can now be
calculated without knowing the function F, by computing the AEf as functions of the Ef
and of the phases, and then averaging over all phases. Such averaging replaces the
random phase hypothesis for our case.

4. The H-Theorem for classical mechanics.

In order to calculate the quantities (25) we substitute on the right-hand side of (20)
the solutions of the unperturbed equation, i.e. we consider the b as constant. Integrating,
we obtain to a first approximation,

(26) bf91 Mw jghj LV f,f',f" b f, bf° 01f, + w f" - of

For a second-order approximation , the b1') are substituted into (20).

b(2)=_ 2 2 Vf,f'.f"V f".f"'.f

i wfwf"
ei(w^•+W1 -+wj;.-w1) t -e

b fbtb
(wJ-+wf" -w1)
 25

Ann. d. Physik , 3, 1055- 1101 , 1929 25

On taking the mean over the phases of the bc, the expression which we then have to
construct will be non-zero only when it only contains terms in the b, i.e. when it contains
only products of the form b ffhJb j:e-b _j. But since b f) only appears multiplied by b° J , we
need only calculate those terms for which

or
f' = f, fns=-fiv

fni fi = - fiv
= f,
or
fly = f,

f'
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

But in the first case, the coefficient Vf,rJ• vanishes: f" = 0 because of (17). The VJrc
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

vanish , as is easily seen by taking e.g. example (16). (It should be noted that in all these
formulae, f stands for the vector f g h j .) The other two options produce:

it + e`(Wf-°'f.-°'r
(27) bi2i _ 4 e I bc, I2 ), -1
Jb i T
f M WIWf• [Wf,+wf"-WI (Wf,+cuf"-Wf)2

Because of (12'), by taking the mean over the phases, i.e. by striking out all terms that
do not contain just terms with the b, we obtain

dE f = MW j (bl(l)b_I(l) + b j b_12) + b112) b°I )

2
4 VJKfIf"A 1-cos(W1.+Wf "-Wf)TE E
M3I^ W4.Wf (Wf.+WI"-Wf)2 I' f "
2
(28) 8 VJV J,I',J" 1-cos(Wf +WI"-Wf)TE E ,
M3 W j. W,W f" (O)f, + Wl" - W f )2 J J

I 2

M3 ^j W2 W2 W2" WJ[ WJEI EJ" WI EIE "

I
J
1-cos ((+Wf" -Wf)T
-W"EE.]-
t J J (Wf,+ Wf"-Wf)2
Also,

ARf2 =M2W4(bjb(lf+bf)b°f)2.

Analogously to the above method we enumerate the number of ways in which pairs of
indices with opposite sign can be set equal, and we obtain,

Iv8 "I +Wf"

J,J',J
2
1-cos(W1, -Wf)T
(29
1
) dE f M3 O )f2, Wf„ ((Of, +Wf" - O )f), EJEJ,EI".
 26

26 On the Kinetic Theory of Thermal Conduction in Crystals

In the same way, we finally obtain for f *f*,

1-cos(Wf,+Wf. -cuf)T
8 IV f,f1f'I
„_-- EE.E. f f f
M3 WfWf.Wf' (a)p +Wf. -Wf)2
(30) 2
1-cos(Wf+Wf. +Wf.)T
EfEf.Ef. .
Wfcof.Wf. (Wf+Wf.2
2 +Wf,)

We thus see that the changes in the Ef produced by the perturbation energy are in general
periodic in time . Only when the resonance condition is of the form
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

(31) Wf+Wf,+ Wf-=0

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

will the relevant terms grow linearly with the time and there exist "transitions" of the
energy between the different vibrations . It is seen that there are then always three
vibrations that participate. In fact , since the products of two Ef appear in (28), there will
always be two vibrations that are excited . The meaning of the first term in (28) is that
anew vibrational state f is generated by the two vibrations f' and f". The second term
corresponds to the decay of the waves f and f" provided the w are positive , if one looks
at the corresponding terms in the expressions for dE f. and dE f.. . If however (of, , for
example , is negative , it will turn out that two waves are at first excited , but that one of
them decays, and the energy of the other will increase correspondingly , simultaneously
producing a scattered wave. For each solution of (31) there will always exist three such
processes, which exactly compensate each other when the system is in thermal equilibrium.
But evidently this equilibrium can only be attained when there exist solutions of (17) and
(31). In fact , as was seen in section 2, there must always exist certain solutions for which
the right hand side of one of the equations ( 17) is equal to ±2,r , and not to 0. The
question , under what circumstances such cases arise , will be left until later . If there are
such solutions , then there will appear resonance denominators on the right-hand sides of
equations (28) to ( 30), compared with which we can consider the V and the Ef as slowly
varyingl . Replacing the summation by an integration (in the limit of infinitely large IV),
we can introduce in the usual way f g' and o)f as integration variables , in place of f' g' h'.
The integration over (of can then be extended to infinity, since the resonance term

1-cos(Wf. +Wf.-Wf)t

((Of, +Wf" -Wf)2

drops off very rapidly . Elsewhere , c of will be replaced by its value at the resonance.
Carrying out the integration will then produce a factor proportional to T , so that finally,
(32) AEf - •(W E Ef"• -io E E E E ••)
f- fC_ f,f.f f f" fff f
f f
where Cf pj is symmetric in the three indices, and is only non-zero when ( 17) as well as
(31) are satisfied. Correspondingly,
(33) AMEJ = 2 JW f2C_.f,f,,f'EfE fEr

I It might look here a doubtful assumption in the classical calculation that the Ef are slowly-varying.
A justification for this assumption is not really called for, since the basis for it will be demonstrated for the
much more stringent approach in quantum mechanics.
 27

Ann. d. Physik, 3, 1055-1101 , 1929 27

The situation is somewhat different for 4 E14 E, . , since f, g', h', in (30) are already
fixed through ( 17) [in ( 17) one will of course have to write f*, g*, h* instead off", g", h'J,
so that one only needs to sum over j'. The integration over the resonance factor should
on]- a2W
y be done after the substitution into (24) It must also be assumed that aE faEf•
is a slowly varying function off . But this does not impose a significant restriction on the
distribution function.
We can now retroactively justify our having retained only the quadratic terms in
the AEI in (23). For the calculation of the higher-order mean values (25), one has to form
products of more than two quantities of the form (26); but after that there appear no
further terms that are proportional to r, only terms of second order in z. If therefore the
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

time ,r is small compared with the "life time" of such a state, such terms can certainly be
neglected. On the other hand one has make sure that one does not choose r to be
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

arbitrarily small, since the resonance will become less and less sharp, the smaller r. This
point, which is unimportant here, will turn out to be of great significance when we
consider thermal conductivity.
After using (32) and (33), and performing the integration for AEIAEI• , Eq.(24)
becomes:
(34)
AW = ITP
^wfCJ,P,f_ aEd I [EIEI.EI*(aEI +-_w1. + aEI. (of-)].
dEf, wI

It follows immediately that the Maxwell distribution

(35) We = A • e

is stationary, since
aW = aW = aW
aEI aE1 aEI.

there, and the expression in brackets in (34) vanishes because of (31). It remains to show
that (35) is the only distribution which is stationary. For this we consider

(36') H=-f ...f W1nWdE.

Its change with time is, from (34),

AH = - f ... f AW(lnW + 1) dE.

But, because of the definition of W,

f ... f AWdE=O,
which could also be verified directly from (24), and

- t =f ... f In W J w1CI,I',I" • a' [EIEI EI,. (w1 + wI, W + w pm J _)1dE.

I I I' I
 28

28 On the Kinetic Theory of Thermal Conduction in Crystals

Integrating by parts and taking the symmetry of Cf r r into account, we obtain

(37) _ ...J1
w
J dE.
3J -WCf,f',f"Ef (Mf dE!
+ wr dE , + wr d
1 f/

But the coefficients C are all positive , and therefore AN is always a 0. When W is
stationary , the equality sign must hold and this is only possible when
dW
(38) w + w dW +w , dW =0,
f dEf r dEf, !dEf„
for all solutions of (17) and (31)
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

Eq. (38) is a system of homogeneous linear equations in the . If we now put

aW
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

dW f
= x f , we obtain , the equations
qEf
(38') xf +xf,+xp =0.

If (17) had all zeros , these equations would have as solution xf = f, xf = g, or xf = h,

respectively, in addition to xf = wf . The processes that matter are those for which the
spatial interference condition mod 2ic is satisfied . They can be described as follows,
provided the f are restricted to -n s f s r : two waves decay for which f + f > x, g +
g' > ,c, h + h' > n, and the wave f + f- 2,c, g + g', h + h ' is formed , with arbitrary f.
Such processes will be referred as "Umklapp processes".
Evidently , (38) has another solution as well, when the xf occur in only one of these
equations, i.e. when there exists only one solution of (17 ) and (31 ) for given f g h j.
Because of the identity E f,_g _h -j = Ef,g,h j one also has to take into account those solutions
for which ( 17) and (31) are satisfied with -f, -g, -h, -j.
If one considers f, g, h, to be continuous, then one can consider (38) as a functional
equation. It will be convenient to define x (f g h j) for all -co < f < +co, etc ., subject to
the condition that x(f +2n , g h j) = x (f g h j). Let as now look at the function x(f) + x(f')
+ x(f"), which will have to vanish at all positions for which ( 17) and (31) are valid. Since
it is also separately the sum of three functions of f, f, f", it will be of the form

a(wf+(of•+wf..)+$(f + f'+ f")+ y(g+g'+g")+ S(h+h'+h").

It must also vanish for Umklapp processes , and therefore be of the form

a(wf+wf• +wf.)+$(f+ f'+ f"+ 2x)+y(g+g'+g")+S(h+h'+h").

Both can only be true when $ = 0. Similarly , y and 8 have to vanish.

dW
For a stationary distribution , therefore, we certainly have , for all
df = dEf ,
dE
f and f, with the above assumption that there is a multiplicity of solutions of (17) and
(31). With this, we have initially only proved that W is a function of the total energy
1: Efg hj . The particular form (35) is only obtained when the system is made to interact
with a "heat bath ". Since these results are sufficiently well known , they will not be
discussed here.
 29

Ann. d. Physik, 8, 1065- 1101, 1929 29

5. The H-Theorem, quantum -mechanical treatment

For the quantum-mechanical treatment of the problem, the variables bf turn out
to be inconvenient, because they are connected with the coordinates by means of a time-
dependent transformation. We shall therefore put

(39) ,j = b fg,Aje'0'r..a.i' ,
af,e.k
so that, from (11),

1
af,a,a,j = (4fek j + f.v k f )
'Wf,s,k,,i
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

It is easy to see that the associated momentum is

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

(40)
PaI..ad = - i M cef,gJ+>,i
One then obtains the equations of motion (without the cubic terms),

(41) M w f,a,),iaf, U,Aj ,

l afa,n.J = z w1,g,Ajaf,e,hJ •

On going over to matrices , one has to note that the a are not real, so that their matrices
are not hermitic . Instead, one has to stipulate that the matrix associated with the
complex conjugate a_ f _8,x_1 = a f.e,h,J is given by

(af*)sk =(af)ki

Added to that comes the commutation relation , which now reads , because of (40),

h
(42) a fa_ f - a_ fa f =
2zMwf.a,k.j
Finally, the expression for the energy will have to be symmetrised, i.e.

(41') Ef = 2 Mwf(af a-f +a_ f af).

From (41), (41 ') and (42 ) one obtains in the usual way,

(a f),m, = 0, when n'# n -1,

(43)
2n >

Ef,n = h (n+2)wf.
 30

30 On the Kinetic Theory of Thermal Conduction in Crystals

Here, n a 0 for wf a 0 and n < 0 for wf < 0 . For the description of the state of our system
all the quantum numbers nfghf will have to be specified. But since E f,_9,_ha = Effhj, one
could really let one of the indices only go from 0 to a; nevertheless it is easier for formal
reasons not to make this requirement, and to define, instead, that

(44) n"-f•-9,-h,-j -(n fghj + 1) .

One will now have to introduce a probability amplitude a(nf) which depends on all
the nfghj . It is constant as long as the cubic terms are neglected, and will change only
slowly if that perturbation is taken into account. We haver
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

2
= tai h
-E )ta(n )
6:(n')
f h IWn n eE"
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

>> f

The W,,,,, are the matrix elements of the perturbation energy. This is, from ( 18) and (39),

g Vff f,.afaf,af„ .
According to (43 ), the matrix elements are therefore equal to zero whenever not all n' are
equal to their corresponding n, apart from the three indices for which

n' =nf -1 ; n'', =nf. -1; nf- =nf. -1.

Taking (43) into account, we thus obtain,

() tai(Th 13/2 ^^(nf+1)(nf.+1)(nf_+l)

a n= h
(45) I Vf,f f (Of WfIw f-
.ei(Mf+w1. +0)1.)t
.a(nf+1,nf,+1,nf..+1).

To a first approximation we shall again treat the a on the right-hand side of (45) as
constant, and obtain

(Th 13/2 y f of + 1) (n f. + 1) (n f.. + 1)

a(1)
(n f,)
t = 4a h ff I wfwf,wf
(46)
ei(wf +Wr-f.) -1
wf + wf• + O )fl •a (nf+1,nf•+1,nf..+1).
f^

For the second-order approximation we shall have to insert the values (46) into (45).
Again, we consider the phases as independent, but in contrast to the classical case they
are not the phases of the elastic waves but those of the a(n), which have no direct physical
meaning . Finally then,

1 Cf. e.g. P.A.M. Dirac, Proc. Roy. Soc. 112, 661, 1926.
 31

Ann. d. Physik, 3, 1055-1101, 1929 31

"12
k(nf'r)I2 = W(nf, =) =, MW ^IVA'1
(47) (nf+1)(nf•+1)(nf"+1) 1-cos(wf+wf,+wf")t
(ppf+(of,+wf")2
wfwf•wf"
[W(nf+l,nf.+l,nf"+1)-W(n))

The of are not written down explicitly and are supposed to be equal on both sides of the
equation. It would appear that (47) only applies to those transitions in which three
vibrations each lose a single quantum, but that is due to our notation (44). (47) leads
again to the condition (31) for resonance, and thus at least one of the w is always negative.
In (47), the integration over the resonance factor cannot be carried as simply as in
(28), for the of are integers and are therefore certainly not continuous functions off. This
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

is connected with a result that has been discussed in more detail by Paulil. For it is not
possible to indicate exactly what the excitation energy is for a single vibration f g h j; in
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

order to measure the frequency sufficiently accurately to be able to distinguish between

neighbouring f s, one needs an observation time T of such duration that

T•dw=T !M •Af=1.

do
But from (7) we have df = , and since is the group velocity, T is of the same
a ff
order as the time needed by a wave packet to pass through the whole portion of the
crystal. During such a long interval, an appreciable number of transitions will have taken
place according to (47), so that the of will also have undergone appreciable changes. f can
therefore be defined only to within an uncertainty of order 1 IT, where T is the life time
of the corresponding state. If we subdivide the f g h j - spectrum into `elementary ranges'
of width 2b, denoting them by FGH , then we can only determine

F+b O+b H+b

NFOHJ afgk
f=F-b g=d-b A=H-b

We now have to assume that at time zero the probability for a particular combination of
the of depends only on the NF . Apart from that , we calculate instead of (47) the sum of

the values of all combinations of the of for which Y n f = NF , since this is the only

quantity that is measurable . To the state NF there belongs the statistical weight

(48) YN = I\
^(NF+r-1

where r = (Nb/2,c) 3 (which should actually carry an index F) is the number of `places' in
the elementary range FGHj . This number has to be chosen such that b(dw/df) - 1/T.
One can now first sum over the of in (47) , obtaining yN(NF + r)/T in place of of + 1, where
the y belongs with the values NF , and not the values NF + 1, etc . Because of (48), this
can also be written yFN+1(NF + 1) /T . Now the resonance factor can be summed over as
before , giving

1 W. Pauli, `Probleme der modernen Physik' (ed. P.Debye , `Sommerfeld-Festschrift'), p.30, Leipzig,
1929.
 32

U On the Kinetic Theory of Thermal Conduction in Crystals

r
dW = h
,r ID FF'F"
rFrFrF-
Y(NF + 1, NF, + 1, NF„ + 1)

WF Wy Wp (NF + 1) (NF. + 1) (NF- + 1)

[W(NF + 1, NF. + 1, NF. + 1) -W(N)] .
Here, DFFF• differs from CFFF" (introduced in (32) ) by the factor

(49) 9 (rF rF/3 + rF, r%11 +... ),

which expresses the increase in the transition probabilities caused by the width of the
elementary ranges . Because of the definition (44), one can reverse the sign of the three
index triplets, and write more explicitly,
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

(50) 4W=2^ prFrWFWF'WFr{(NF+1)(NF,+1)(NF-+1)

y(NF +1, NF, + I, NF•, +1) • [W(NF +1, NF, + I, NF. + 1)-W(NF)]
+NF NF Ny(N)[W(NF -1, Ny -1, NF•. -1)-W(N)] } .
In the limit of large quantum numbers, (50) tends to (34), apart from y and the factor
(49), as it should. Obviously, (35) is stationary here, too, since it follows from (31) and
(43) that
Wo(NF, NF, NF.) = Wo(NF + 1, NF, + 1, NF. + 1).
Admittedly , this relation is not rigorously valid , since (31) only holds to within an error
of order b (&o/df), because of the width of the elementary range. But that is the only
accuracy to which one can define (35) at all, since the energy is just determined to within
h/i . Obviously it makes no sense to require exact stationary behaviour here.
Now consider the quantity

(51) H = W(N) lnW(N)

and its change with time (renaming the summation indices in the first term),

AH=-I:1nW•4W
_ hT DF, F•,YWFW, WF" NF NF
NF. Y(N... )
2^ rFrF' rP
N F FF"
[W(NF, NF, NF,) -W(NF -1, NF. -1, NF, -1)]
Jl nW(N) -1nW(NF -1,...)]
Since In x is monotonic, we always have AH a 0 [because of (44), WFNF is always a 0], and
the equality can only hold if we have

W(NF -1, NF, -1, NF. -1) = W(NF, NF', NF--)

for all the F F' F" for which (31) and ( 17) are valid,
It follows from this condition , making the same assumptions as in sect . 4, that

W = W(Y WF NF)
We shall now have to investigate under what circumstances these assumptions are valid.
 33

Ann. d. Physik, 3, 1055- 1101 , 1929 33

6. Discussion of the conditions for interference.

The question suggests itself in how far the conditions for the H-Theorem are
already satisfied for the simple case of a linear lattice with the dispersion law (10).
Because of (17) and (31),

sin2sin 2 =±sinf 21,

i.e. either

sinf 4 f cos f cos f4 = o,

or
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

sinf 4 f ' sin-sin-=0.

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

Each of these conditions can only be satisfied when one of the three variables f, f', f"
vanishes. But in that case , as has been seen, the coefficients Vffro vanish . This does not by
itself imply that the corresponding transitions do not take place , for in (28) a factor wf2
occurs in the denominator which compensates for the vanishing of the numerator. The
reason for this is that , because of ( 15), the vibrations with small w have a very large
amplitude for the same excitation potential . But, for these transitions , f and f are
neighbouring vibrations and the energy flux only changes by an amount which is of the
order of the resonance broadening . They are therefore of no importance for the thermal
conductivity . They are also not sufficient for the proof of the H-Theorem ; for if the
resonance were to be sharp, then each of them would exist in only one such process, and
other stationary states would then exist (see p . 14). Because of the resonance broadening,
f will be present in several of the transitions , but their number is always finite: while an
equilibrium could be established in this way, this could only happen in a time interval
which is proportional to N. In reality , there are other reasons for the establishment of the
equilibrium : our calculations were only taken to second order , so far , whereas for the next-
higher order of approximation there appears an interference condition which is of the form

wf+Wp * wf-+wr =0,

with

f+f,+f„+f,,={+
2x

We now have one more variable, and the equations will always have non -trivial solutions.
One is led to the same conditions if one takes the series expansion of the potential function
to terms of fourth order in the atomic distance . But for such processes, there must always
be three oscillations which are excited , and the decay intensity of a given wave is
proportional to the excitation of two others. Since all oscillations have an average energy
kT for high temperatures , the linear model shows a thermal conductivity which is
inversely proportional to the square of the temperature.
Against that , an energy exchange takes place between neighbouring frequencies
already for second-order processes , and this fact can be used to explain the absorption of
the residual radiation . For this purpose, it is quite permissible to restrict oneself to the
linear model'.

1 W. Pauli, Verh. d. Dtsch. Phys. Ges. (3) 6, 10, 1925

 34

34 On the Kinetic Theory of Thermal Conduction in Crystals

A different situation presents itself however when one goes to more than one
dimensions . Let as discuss the two- dimensional case . In space, circumstances are
certainly more favourable than in the plane and if it can be shown that there are always
solutions of (17) and (31) in the two- dimensional model, then the same applies a fortiori
to three dimensions.
Compared with the linear model we now have six variables, f, f, f", g, g, g" , and
three equations , i.e. two from (17) and one from (31). In addition, we also have the
flexibility of being able to vary the index j for each of the three waves, but for the
moment we shall not make use of this, and put j = j'= j"
One can now think of a surface which is constructed in such a way that one plots
the values off g in a plane, and the corresponding value of w in a direction normal to the
plane . Specifically, we use the same assumption about the forces that was used by Born
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

and v. Karman, l.c. (only neighbours at distances a and aN/2 are taken into account). In
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

that case one knows that the intersection of such a surface with the plane g = 0 is a curve
of type (10). In order to construct the solutions of (17) and (31) for fixed f and g, one
would have to shift the surface in the fg-plane by the vector +f, +g and normal to it by
wf9. The points of intersection of the shifted and unshifted surfaces produce the desired
values of f', g'. If one now thinks of the whole graph as extended with period 2,c, one
obtains at the same time the solutions that correspond to the lower of the equations (17).
Fig. 1 illustrates the intersection of the graph with the plane g = 0, and the shifted curve
is constructed on the assumption that g = 0 for the first wave. In agreement with the
results on the previous page, the curves then have only those points in common which
correspond to the trivial solutions. But the surfaces, too, have no other points in common,
since the width of the graph in the g-direction (and in any other direction of the fg-plane)
is zero at the origin. If on the other hand g 0 for the first wave, the situation is exactly
the same, since again the trivial solution f" = g" = 0 exists, and therefore the origin of the
unshifted surface will again lie on the shifted surface. It is easy to visualize these
relationships by observing that the surface in the neighbourhood of the origin is similar
to a double cone about the w-axis, but which is distorted in the f- and g-directions.

Fig. 1
Graphical solution of the interference conditions
 35

Ann. d. Physik, 3, 1055-1101, 1929 35

Nevertheless, points of intersection could exist that lie very far from the origin of
both surfaces. This would depend on the specific dispersion law. However, such roots
would not exist for arbitrary values of f and g, so that they would not serve our purpose.
But there exist two such w-surfaces that correspond to the two values of j (in two
dimensions). Ifj, j', j" are now chosen to be distinct, then the origin of the shifted surface
will no longer lie on the unshifted surface, or that of the unshifted surface on the shifted
surface. This means that the surfaces possess at least one curve of intersection. There
therefore always exists a multiplicity of solutions of the interference conditions. The
processes that matter are those for which the waves have different states of polarization.
It is also evident that, for sufficiently large f, g, there will always exist points of
intersection which lie outside the range -sr < f < x and which therefore belong to
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

`Umklapp' processes.
Finally, it might be conceivable that all coefficients Vffr could vanish for just such
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

solutions. This is impossible for the following reasons : according to the calculations by
Born and v. KArman, the dispersion law contains two matter constants , even in the most
idealised case , i.e. it contains at least one dimensionless constant which irHluences the form
of the law and, through it, the solutions of the interference conditions. The perturbation
potential which determines the Vffr equally depends on at least two constants, which
however have nothing to do with the former since they are obtained from higher-order
derivatives of the force law. At best, therefore, it might just be possible (even if highly
unlikely) that the coefficients in question vanish for a special choice of these constants;
but for such a fortuitous relation to occur in nature on a regular basis , is hard to believe.
We therefore see that for a three- dimensional crystal lattice the conditions are
satisfied for the existence of the H-Theorem and for a finite value of the thermal
conductivity.

7. Classical formulation of thermal conduction.

Let us now consider those processes which depend on the location in the crystal.
For this, a Fourier analysis with strictly monochromatic waves (such as had been used in
the previous sections) would be unsuited, since space-dependent amplitudes of the waves
can only be produced when more than one frequency is available. Correspondingly, one
will have to proceed as described below, using again the linear model. The generalisation
to three dimensions is straightforward.
We divide the chain of N atoms into NIL sections of L atoms each, and divide the
spectral interval -x < f < it into L parts of length 2x/L. One can then investigate the
intensity of the vibrations in the interval f of the v-th section of the crystal.
Correspondingly we define,

L
(52) Bf,, = N jbf+v/L.e'v

Here, xf is an arbitrary number between 0 and 1. The inversion of (52) gives,

r-K )v
1 (52') bf+r/L = BfV e-:(
FN7
To visualize the meaning of the quantities Bf , we shall assume that only a single Bf,, is
excited, and that all the others are zero. From (52'), (11) and (6) it then follows that

L ij(f+^IL)(n-Z+.fL)+ wf+.1Ll•t
un-1rl-Bf'
c
 36

36 On the Kinetic Theory of Thermal Conduction in Crystals

Neglecting (1/L)2, we can write

W1"IL=WJ+LvJ

where v1= dw1/af is the group velocity (measured in terml of the atomic distance). The
sum over 9) can then be performed aid we obtain, to within a constant factor,
(53) u,, = Bfv R(n-vL+x1L-vft),
with
R(x) = sinax
nx
For each f we therefore have a series of NIL wave groups , which propagate at distances
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

L from each other, one behind the other. xJ determines their position at time 0. If e.g.
xJ = 0, they fall exactly into the above-defined sections of the chain. In general,
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

corresponding to (53), the amplitude of the f-th vibration in the v-th section will be
denoted by the quantity
(54) Af, Bfv.R(vL-v'L+KIL-vft).

The perturbation potential then becomes, expressed as a function of the B,

V7 =1
3 L , ffr e
(55) MI 1 1f
+L J+L J. L J-)BJvBJv„BI,.v,,

Here we have first of all neglected the change in Vfff' within a spectral interval. Because
of (17), we can then sum only over values for which q, + qp' + gyp" = 0. Furthermore, the
terms containing the of are associated with the overlapping of the waves. Effects caused
by this are evidently small and of order V (1 /L) and can therefore be neglected. Summing
over op, q', we are left with

e-tq^oJ+mJ'+'r)
V= 1jV111.
3 L
(56)
R(vL-v'L-x1L+xr.L)
R(vL-v"L-x .L+x1'.L)BfvBfv Bf"1p

Later we shall average over all the x1. There will then appear in (56) not only terms with
v = v' = v", but also those with neighbouring v's. This would have to be taken into
account for an exact calculation, and would lead to the conclusion that the B1,, for
arbitrarily distant v can no longer be treated as statistically independent. But we would
not destroy any essential property of the problem if we were to consider the B in (56) as
slowly varying functions of v; we can then replace v' and v" by v, because of the rapidly
decaying factor B. When this is done, (56) will then assume the form (18), apart from the
factor V (NIL); only the meanings assigned to the B and the f will be different. We can
therefore use Eq. (34) for the distribution function of the e1,. = B1,,B_1W12. However we are
not interested in the excitation potential of the B, but rather in the vibrations in a given
section, i.e. in the quantities (54).
 37

Ann. d. Physik, 8, 1065- 1101, 1929 37

The change with time of the A consists in the first place of the explicit time-
dependence of (54), which is already present in the case of harmonic forces and
corresponds to the propagation of the wave packets, and in the second place, of the
scattering processes. If we again neglect the product of the intensities of these two
processes, i.e. quantities of order V.(1/L), then we can deal with these two changes
separately. In the case of the scattering process we have to put t = 0 in (54), and it
follows that (34) is also valid for the distribution of the

(57) Efv = wwAf,,A_ j,,

Also, another change takes place according to (54). The time ti is small compared
with L/vf - which must obviously be true, if one is to be allowed at all to consider the
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

change in the state as small . But then

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

d'Af,, =ryBf vfR'(vL-v'L+xfL)

and

dEf, =zwfI Bf,,.Bf,,.vf[R(vL-v'L+ipfL)R'(vL-v"L+xfL)

V V.

+R(vL-v"L+xfL)R'(vL-v'L+xfL)J .

We now average, first of all over the phases of Bf,,, and next over the xf from 0 to 1,.1
What is left, is

Ef" 2LEf^ vf[R4(v-v'+1)- R4(v-v )J

A

= 2L vf(Ef,r+i -Ef,)

The higher averages dE f,,4E f„ , etc., do not make any contributions that are

proportional to v. Therefore, according to (24),

dW =-2L ^aif-[vf(Ef,' +1 - Ef,)WJ •

This becomes, because Evf =0,

f

(58) dW=-2L1vf(Ef,,,+1-E1,,) E .
fv f^

In a stationary state, (34) and (58) will cancel exactly.

1 If one were to set all xf = 0 , then B' = 0. It thus quite pointless to require that at time zero the two
wave packets belonging to f should happen to coincide: this would never occur again, because of the
different group velocities.
 38

38 On the Kinetic Theory of Thermal Conduction in Crystals

Let us now assume a position-dependent temperature T = To + p v and write

accordingly,

W=11 Wv(EIV)
V
(59)
WW=Wo•[1+pvaldT°+pA(Ej)l

Here, Wo is the Boltzmann distribution (35), A is a function that is no longer dependent

on v and has to be so chosen that the whole expression is stationary. The function that
is thus obtained will depend differently on Efj, and Ef -j,, , i.e. on waves moving to the
right and left, respectively. (The situation is analogous to that in the [kinetic] theory of
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

gases : at a point where a temperature gradient exists, more atoms of a given energy will
be moving in one direction than in the other.) A non-zero average energy flux will
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

therefore be associated with the distribution (59).

We shall calculate, first, the expression (58) for the distribution (59), up to terms
of order p2,
1
-2LWojvvf(Ef,v+i Ef,v) k2
f,v
By means of a "partial sum " ( interchanging v and v+l in the first term), this can also be
written in the form,
p Wo2IvfEr[(v-1)-v]= PT W21vfEfv
(60)
2L U f, 2L kT f,
first two terms will drop out, since
Substituting (59) into (34), the contributions from the
the distributions Wo and aW0/oT are stationary with respect to scattering processes.
For the A- term one obtains,

AW=pWy, 10)fc. f"

v f f,.r
(61)
[EfEpEf v(cof
aEf +O
aEf•
+w,- aEf.
af
By combining (60) and (61) we obtain the equilibrium condition,

(62) 2L k 1 yvfEf = Ywf cflf'

f f,f1"
dEf [L fEf,Ef (cif +a^f aEf. +cif .dEfN

I have not been able to solve this partial differential equation for A. The function A does
not have a simple form, since ( 62) is not separable in the different Ef .
One can, however, first ask oneself whether this equation possesses a solution at all.
To this end we note that the differential expression on the right-hand side is self-adjoint.
In order for the equation to have a solution , it is necessary that the expression on the left-
hand side of (62 ) should be orthogonal to the solutions of the corresponding homogeneous
equation . But, as was seen earlier, such a solution of the homogeneous equation is a
 39

Ann. d. Physik, 8, 1055-1101, 1929 39

function of YEr , and in that case the contributions of the waves travelling to the right
r
and to the left will cancel out, on integration. The reason for this is that the total energy
of all waves in the v-th section will remain unchanged. Energy will be transported in the
direction of lower temperatures; nevertheless, on average the same amount of energy will
be brought in from the other side. If an additional solution of the homogeneous equation
were to exist which depended on Erb and Ef -j in distinct ways, then (62) would have no
solution with finite R, i.e. the thermal conductivity would become infinitely large. This
is, e.g ., true in the case of the continuum.
It is also possible to determine the temperature dependence of the thermal
conductivity without having to know the solution of (62). For this, let us put
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

(63) kT
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

In terms of these variables, (62) becomes

E wf Cff r axr (xrxr•xr"(wr a f^ + wr- x^ + (Or ^. )l

r^r',r"
2:r
2 LkT 2 vrxr
A will thus become proportional to I/T2 . On the other hand, the energy flux becomes,
from ( 14), (35) and (59),

(64) S =KT p f ... f vrx1Wo (x).(x) dxr dxr.... ,

where the constant in (35) is normalised in such a way that

f...fW(x)dxrdxr ...= 1.

It follows that S is proportional to 1/T , in agreement with Debye (l.c.) and with
experiments.

8. Quantum -mechanical approach to thermal conduction.

For low temperatures the effect of quantization makes itself felt, and the dependence on
T will then become quite different. The calculation is analogous to the calculations in
sect. 5, with the only difference that the variables (52) and (54), resp., have to be used.
It should be noted that, according to p.31, we must choose 'r << L/vr since otherwise the
change in the system can no longer be treated as small . The `life time' which was
introduced and determined there by the intensity of the scattering processes, i.e. the
magnitude of the Cfff- , has now acquired the meaning of L/vr . This determines at the
same time the width of the resonance, of order , so that the `elementary
L = L f
regions' will extend over a number of the spectral intervals f - i/L, f + ,r/L. With these

1 A. Eucken, Ann.d.Phys. 34, 185, 1911.

 40

40 On the Kinetic Theory of Thermal Conduction in Crystals

restrictions in mind the arguments of the previous section can now be applied. We
therefore have to use (50) for the changes in the distribution caused by the scattering
processes.
In addition, we shall have to take into account the spilling over of waves from one
section of the chain to the next. The quantum-mechanical analogue to (58) could be
derived by using perturbation theory. Consider the phases of the probability amplitude
a(nr), where the nf,, denote the energies of the Af,, ; i.e. the n1,, are the indices in the
matrix formulation in which the Ef,, in (57) become diagonal. If we were to look upon the
a(n1,,) as independent, then the change in W would become proportional to t2, for small
times. The reason for this lies in the fact that one has started from a special initial state
(cf. footnote on p. 23). One would therefore have to make a transformation from this
system to one in which the intensities Bf, are diagonal, then average over the phases, and
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

finally, after transforming back, average over the ic1. But it is not necessary to carry out
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

such a calculation, for one can argue as follows: for sufficiently small time intervals one
certainly needs to consider only those changes of state for which a single quantum from
one section is transferred to a neighbouring sections. On the other hand, the change in
W over longer time intervals is known if the cubic terms are ignored. For times large
compared with L/vf, W can only be noticeably different from zero for those values of n1,
which correspond to a wave packet travelling with the group velocity, if one neglects
dv
terms of order which represent the "melting away" of the wave packets. Thus,
L2 a ff
for long times,

(65) W(n1,,,,t ) W(n1^-vfL,O).

On the other hand, for small times,

4W _Zp,,k[W(n,...,nf,,s+l,nf,,,+l,...)
(66)
V,f
-W(n,...,nf,v,-1, ni, ,...)].

One may compare (65) and (66), even though in (66) averaging over the phases was carried
out, i. e. an observation took place . This is because it was carried out in a matrix
representation in which the energy was diagonal (since the B1,, are constant in time
without the cubic terms ), so that when a measurement has been repeated, it will always
give the same result. (66) will therefore lead to (65) when applied successively over smalll
time intervals - and this is only possible when the factor pf,,, = nf,,,(v1/L). Therefore, (50)
immediately becomes

1 In the perturbation calculation , this situation is represented by the form (43) of the matrix element.
It was wrongly interpreted by v. Ki rman (see Introduction ), who thought that the energy exchange could
only take place between those sections whose quantum numbers differ by 1. In fact , it is the quantum
numbers of a single section , before and after a transition , that have to differ by 1.
 41

Ann. d. Physik, 8, 1055- 1101 , 1929 41

v
(67) -^nf,, L [W(...,nf,,^1+1,nf,,,+1,...) -W(...,nf,,^l,nf,,,,...)] .
v,f

We now use again the trial function (59), and obtain the condition for a stationary state,

h
(NF +1) (NF +1) (NF' +1)
2x^
^^1
F,. r1 F ^F ^F ^F" {
(68) 7(NF +1A'F +I, Ni_+I) [ NF +1, NF +1, NF. +1)-A(NF)]
+NFN,VNl-Y(.LV)LAkNF -1, NF -1, N -1) -Ak1V)] }

T NF
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

Of course , (68) is still more difficult to assess than (62), since it contains (62) as a special
case . What is of great importance for our ability to solve (68), is the previously discussed
lack of sharpness of the resonance broadening . In its absence , (68) could be separated into
a series of systems of equations , each belonging to a value of the energy; the same total
energy would then correspond to the two combinations of the values of NF which appear
on the left-hand side of ( 68). Let us now consider the systems of equations whose energy
is smaller than hcu (n/2)/(2n), because of (31). Evidently all indices NFL, with F,G,H a n/2
would then have to vanish , and therefore no Umklapp processes could exist in the sense
of p. 28 . In this case it is easy to see that the rank of the inhomogeneous system of
equations is greater than that of the homogeneous ones, so that there exist no solutions
at all . But in reality such a separation is an impossibility, since processes will occur for
which the energy changes by the resonance width - even though such processes cancel
out on average. Again , as might appear reasonable , one cannot argue that the resonance
broadening becomes smaller with increasing L. The width 2n/L of the spectral range will
then also decrease , so that this resonance broadening will certainly extend over several
such intervals, as mentioned before.
The conditions for solving ( 68) are exactly the same as those for (62). However,
one cannot specify the temperature dependence of the energy flux without determining
the function A, for the transformation (63) becomes meaningless in this case . Once again,
the situation becomes somewhat more transparent when dealing with the low temperature
limit. For at low temperatures the `Umklapp ' processes , whose importance has already
been demonstrated , become more and more rare since the presence of short waves is
needed for them to occur . It would therefore be reasonable to perform the calculation as
though a complete equilibrium were established, leaving Umklapp processes aside. In other
words , the above-mentioned distribution would imply a maximum entropy with fixed J;
the value of J would then be changed slowly by involving the Umklapp processes.
Analytically, this is expressed by taking the left-hand side of (68), which can be
regarded as an operator acting on the function A, and splitting it into two parts,

(69) D = D° + D* ,

where the operator Do involves all those terms for which no Umklapp processes take place;
or, to put it more precisely , Dc involves all coefficients DFF'F° for which

(69a) F+F'+F"=0.
 42

42 On the Kinetic Theory of Thermal Conduction in Crystals

For sufficiently low temperatures, D* can then be treated as a small perturbation,

and we shall first consider the equation

(70) D°W° = 0.
Its solution is obviously an arbitrary function of

and of
E = J NFh(OFn
J=IF.NF,

since with such a function we have for all solutions of (69a),

Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

W°(NF+1,NE'+1,Nr + 1) = WO (NF, NF,, NF^)

Because of the previously-discussed "resonance width ", a function of E is not an exact

solution but is only approximate , to the degree of approximation used here.
Using a notation that is slightly different from (59 ), we shall choose as a trial
solution,
_E
(71) Wo = e kT O(J)

If we now add as perturbation the operator D*, the equation will not have a solution left
at all that would only differ a little from ( 71): we shall also have to include a temperature
gradient , and look at the wave propagation. For the dependence on position we proceed
similarly to (59) and set
W=W°+Wi+W2,
where Wl is the position -dependent part of (59) and W2 is a position- independent function
which is to be determined in such a way that W is stationary.
Neglecting all higher powers of p and using the fact that D* is also of order p, we
are left with the equation

(72) D°W2 = -D W° - p I I 2n hMkT 2 N

F
F vF . W°
L
The inhomogeneous equation (72) has a solution only when the right-hand side is
orthogonal to all solutions of the homogeneous equation . But we need not require that
it should be orthogonal to arbitrary functions of E, since these are only approximate
solutions of (70), as noted above . The right-hand side should therefore become zero when
it is multiplied by an arbitrary function of J and then summed over all values of NF . We
have here a kind of `degeneracy', since the left -hand side possesses several solutions. We
are therefore forced to look for the particular basis of the solution system (i.e. the
functions 0 in (71 ) ) for which these conditions are satisfied . That this should be possible
at all, is most easily seen in the following way. We note expression (48) for the y and
take into account that the terms in odd N cancel out because of the formal convention
adopted in (44). D*Wo is then of the form

h, D F,F',F° O ) w c° " N N 'r .. + N N ..r ' + N 'N -r 0 (/ J - Oh J ± 2 n

F F F ( F F F F F F F F F) L l ) ( )]
2z rFrF,rF„ F
FF'F
 43

Ann. d. Physik, 3, 1055- 1101, 1929 43

The bars indicate the mean values which are obtained with J kept constant ; they are
therefore are still dependent on J in some way. However, one can restrict oneself to
values of J which are large compared with 2,r , since J already has the value x when a
single lattice vibration of wave length equal to the atomic distance is excited by a single
quantum . One can therefore put

0(J) - 0(J ± 2n) = +2^r d ,

and obtain an expression of the form

(73a) Y'(J).
dO
where Y' is a given, but very complicated , function of J. In the same way one obtains
from the second term on the right- hand side of ( 72) an expression of the form
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

(73b) px(J) • 0(j),

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

and the condition for being able to solve ( 71) is the differential equation obtained by
setting (73a) equal to ( 73b). Its solution is
pjxdJ
O=a•e .
This function , when substituted into (72), then produces that particular solution of the
inhomogeneous equation ( 70) in whose neighbourhood a solution of the exact equation (68)
can be found.
For sufficiently small energy gradients p, the expression ( 71) will only be different
from zero for small J. It is then permissible to expand 0 in a power series in J. This
gives for We the expression
E
Wo =ae x"(1+aJ),
and the energy flux is seen to be proportional to
a•T5
with a proportionality factor that depends neither on p nor on T. It now remains to
establish the dependence of alp on the temperature.
For this , we multiply the right -hand side of (72) with an arbitrary function of J,
e.g. with a-fiJ, and sum over all the NF. The result must vanish, since this was the
condition imposed on the function 0 in the first place . We can restrict ourselves to small
values of J here , since large J become very unlikely , because of the factor e-E/kT. An
expansion of 7 is therefore justified , and we can calculate the coefficient a by dropping
all higher-order terms . The second term gives

ho,,
FrF VP 1 2x!Y
P 2 L luu h-
F 2nkT
e 2akT a 2akT _ 1

Here, we have to take the average in the normal way, i .e. we now no longer have to keep
the value of J fixed. For low temperatures, F can be taken up to 0o as usual , and F/kT
can then be introduced as a variable . It can then be seen that the coefficient of p in the
second term becomes proportional to P. The coefficient in the first term has a more
complicated form. It can be written,
 44

44 On the Kinetic Theory of Thermal Conduction in Crystals

F+ F' F + F'
' aYKFF' ' hm, hm,. ho,,
FF' (e2skT _1)(e2xkT 2xkT 1-e 2AkT

Here one has to make use of the fact that the odd terms in F drop out, since for them
the vibrations with -j exactly cancel those with j.
The dependence of the coefficient KFF' on F and F' cannot be obtained without
some complicated calculations. But at best it could tend towards zero with F like some
power of F. The other factor can be estimated in the following way. We are only dealing
here with the coefficients for Umklapp processes, where F + F' is always > n , so that at
least one of the two F must be greater than ,c/2. The above expression will certainly
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

become greater if it is replaced by

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

1 2nkT
$•
al KFF'•2nkT (F 2 +F 2)
FY kwo / 2) \ 1 F'
1 e 2xkT _ 11

It is therefore certainly of the form

h°'(2)

a•K(P). e 2,rkT

where K(T) behaves like a power of T in the neighbourhood of T = 0. The ratio alp , and
hence the thermal conductivity, therefore grows exponentially for small temperatures, i.e.
roughly like
ie
e2T.

This result was only obtained by using many approximations, yet most of these
appear reasonable. The most dangerous step would appear to have been to cut off the
expansion of the potential function at the cubic terms. For high and not too low
temperatures, this would seem to be legitimate, for the following reason . We had seen in
section 6 that the fourth-power terms for the decay of a wave require the existence of two
other waves; at high temperatures, they will therefore contribute to the thermal resistance,
which is proportional to P. Correspondingly, the higher-order terms produce yet higher
powers of T. But since at high temperatures the thermal resistance is markedly
proportional to T, it can be safely assumed on empirical grounds that these higher-order
terms do not contribute. And just because of this temperature dependence, their relative
effect will become less and less with decreasing temperatures. On the other hand, this
argument is no longer valid for very low temperatures. In that case, the above estimate
for terms of the next-higher order can only be made with exponential terms ee/3T
precisely because of the combined action of three vibrations. For that, however, much
higher powers of T will have made their appearance. It is therefore safe to say that while
in principle the effect of higher-order terms becomes appreciable at sufficiently low
temperatures, nevertheless such temperatures are so low that they play no role in practice.
Giving exact values for these temperatures would require an estimate of the relative orders
of magnitude of individual terms, which is of course not possible as long as one cannot
even give the order of magnitude of the effect of the cubic terms under consideration here.
 45

Ann. d. Physik, 3, 1055- 1101, 1929 45

9. Influence of lattice perturbations.

The assumption up to now has always been that we are dealing with a perfectly regular
lattice. But it is known that in real crystals there exist impurities as well as imperfections
in the periodic structure of the lattices. In what follows, we shall now estimate the effect
that such imperfections have on the thermal conductivity. Here we shall make the
following assumptions about the form that these imperfections take. In a region that
encompasses many atoms but is yet small compared with the crystal as a whole, the
elastic constants are assumed to have values different from those at other locations in the
crystal. This includes the cases where a few foreign atoms are present, as well as cases
where some atoms are absent altogether or do not occupy their correct positions. The
assumption does not apply to boundary layers, where two differently constructed or
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

differently orientated lattices meet: thus, here as well, our discussion does not apply to
micro-crystalline materials.
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

It will be seen later that what is important above all is the behaviour of the long
waves, and we shall therefore assume that the wave length is large compared with the
dimensions of the lattice imperfection. We can also replace the crystal by a continuum
whenever we are dealing with long waves, although we cannot include the case of
anisotropy. All this is just done to simplify the calculation, but makes no difference to
the results that are obtained. In other words, we perform our calculations as though the
frequency for small f, g, h were given by

J2+g2+h2
(74) wJsy = V 2za
The model is exactly the same as that of Debye (l.c.), only the analysis is carried out in
a different manner. If we were dealing with lattice imperfections alone (i.e. static
deformations of the lattice), they would, by themselves, not give rise to a finite thermal
conductivity. For according to Debye, the mean free path will become proportional to
1 /w4 there; but one must not, as was done by Debye, replace w by a mean frequency.
Instead, the contributions of all waves must be considered individually. In Debye's
formula (54) we must interpret l as the free path of a single vibration, c as the specific
heat of this same vibration, and then sum over all vibrations. It is known that
2

c= •k.
2gckT(e2xkT -1)
According to Debye, d is proportional to 1 /w4, and the number of vibrations in the
frequency interval w, w + do is proportional to w2dw , so that we shall finally have to
consider the integral

h
I hto
2 w2dW .

47C2kT2w2 e2xkT -1

But the integrand behaves in the neighbourhood of w = 0 as 1/w2 and the integral
therefore diverges.

1 Cf. e.g. B.Smekal , Z. Phys. 45, 869, 1927

 46

46 On the Kinetic Theory of Thermal Conduction in Crystals

The situation is changed, however, if one combines the lattice imperfections with
anharmonic terms. The only role that the imperfections then have, is to invalidate the
conservation law for the quantity J. In that case they easily contribute a finite amount
to the thermal resistance, whose temperature dependence at low temperatures will now
be determined.
For this we note that with the above assumptions the perturbation energy, taken
as a function of the displacements, is of the form
2

A' dx)
where A is a function of position which is only different from zero in a region having the
above-mentioned property. In the three-dimensional case, the u will have to be considered
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

as vectors and, correspondingly, A as the coefficient matrix, and the second-order

by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

derivative has to be replaced by the operator A.

At any rate, the perturbation energy as a function of the coordinates (39) has the
form

2 2
(75) B afghjapg'hfwf WI' ,

where B is a constant. We have assumed here that the region over which A differs
appreciably from zero is small compared with the two wave lengths

2 2ira
2 and i 27ra , respectively.
f+g 2 +h f 2 + g+h ^2

The perturbation energy matrix is calculated quite analogously to section 5, and the
change in the statistical function W caused by this perturbation will then be of the form

B 1-cos(wf+wf.)r
^= M2 ^wfwf,(nf+1)(nf,+1) w
, 2
( f +wf )
•[W(nf+1,nf,+1)-W(nf,nf')]
The previous considerations also apply with respect to the time dependence and the
resonance broadening. Since the resonance is located at wf =f' , we obtain, analogously
to (50),
AW = const. T I wF2 { (NF + 1)(NF, + 1) y(NF + 1, NF' + 1)

[W(NF + 1, NF' + 1) - W(NF, NF' )] - NFNF' y(NF, NF' )

• [W(NF, NF,) - W(NF -1, NF. -1)] } .
We conclude from Debye's result that the scattering decreases rapidly for small F, so that
the above term can be regarded as small compared with the anharmonic forces for low
temperatures. Accordingly, we introduce this term in (72) next to the term with D*.

Integrating then produces a contribution a • const. I (F2FNF , so that we obtain an

F,F
additional term in the ratio p/a, which is proportional to T6. In other words, in the
thermal resistance term ply we obtain a term proportional to T. This term is certainly of
decisive importance for very low temperatures. It therefore depends on the magnitude of
the lattice perturbations whether the exponential decay of p/y, calculated in the previous
section, is observable at all,
 47

Ann. d. Physik, 3, 1055- 1101 , 1929 47

10. Comparison with experiment.

So far, little can be said about the agreement of our results with experiments. It
was already mentioned that the linear law for T>> ® has been well confirmed.
Experimental results obtained up to now admittedly show that the thermal resistance for
lower temperatures shows a more rapid decreases, but do not yet show whether they agree
for very low temperatures with the relation required here,
1e
KT +A•T".e 2T

At first sight, the results for diamond2 would appear quite baffling: because of the
high characteristic temperature, one obtains very low values for T/®, but the thermal
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

conductivity has been shown to be practically unchanging over this whole region. It is
possible that the reflexion of the waves at the boundary surfaces of the crystal already
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

plays a role there. But its contribution to the thermal resistance would depend on the size
of the crystal, and one could then no longer speak of thermal conduction in a
phenomenological context. This conjecture is reinforced by the fact that the experimental
results are strangely difficult to reproduce. In any case, the theory derived here would
require that an increase in the thermal resistance with increasing temperature should be
found, even for diamonds, provided the temperature is increased up to close to the
characteristic temperature.
11. Summary.

We investigated the mechanism of thermal conduction in the crystal lattice. It

turned out that the condition for the existence of a finite thermal conductivity is identical
with that for the existence of an H-Theorem. We further saw that this condition is
satisfied for a three-dimensional lattice, and investigated which aspect of the dispersion
law is responsible for this. In the course of these investigations it was found that this
mechanism is considerably more complicated than seemed to be the case after Debye's
first attempts. Still, Debye's result concerning the temperature dependence at high
temperatures was verified. By contrast, infinite thermal conductivity was produced when
the atomic structure was neglected. For the physically equally meaningless case of a linear
chain of atoms, the result was proportional to 1 /T2 .
For low temperatures, an exponential decay of the thermal resistance was derived,
which can however be superimposed by a term proportional to T, due to lattice
perturbations.
I am indebted to Prof. Pauli for suggesting this work and for supporting me in the
course of it with much valuable advice. For this, I wish to take this opportunity to
express my most sincere thanks to him.

Zurich, Phys. Inst. d. Eidg. Techn. Hochschule, July 1929

(Received on 24 October 1929)

REP comments:
This paper treats the thermal resistance of crystals following Debye as due to the
anharmonic part of the force between atoms, which causes couplings by which the

1 A. Eucken, Ann. d. Phys . 34, 185, 1911.

2 A. Eucken , Verh. d. Dtsch. Phys. Ges. 13 , 829, 1911.
 48

48 On the Kinetic Theory of Thermal Conduction in Crystals

different modes exchange energy. In this it is important to take account of the

conservation of pseudo-momentum (also called crystal momentum or wave vector) . In
a continuous medium this conservation would be exact; in a crystal it can change by
multiples of the Bragg vector. This process is called in the paper the `Umklapp' process.
Such processes become very rare at low temperatures and the theory therefore predicts
that the thermal conductivity will rise rapidly at low temperatures exponentially with 1 IT.

REP comments:
This paper treats the thermal resistance of crystals following Debye as due to the anharmonic part
of the force between atoms, which causes couplings by which the different modes exchange en-
Selected Scientific Papers of Sir Rudolf Peierls Downloaded from www.worldscientific.com

ergy. In this it is important to take account of the conservation of pseudo-momentum (also called
crystal momentum or wave vector) . In a continuous medium this conservation would be exact; in
by WEIZMANN INSTITUTE OF SCIENCE on 02/21/16. For personal use only.

a crystal it can change by multiples of the Bragg vector. This process is called in the paper the
`Umldapp' process. Such processes become very rare at low temperatures and the theory therefore
predicts that the thermal conductivity will rise rapidly at low temperatures exponentially with 1/T.