,:
•
CLASSICAL STATISTICAL MECHANICS
19:P
. ".
Centro Bras11e1ro de Pesquisas F1s1cas
CONTENTS
~~ .... 
I .
CHAPTER I
STATISTICAL TRERMODYNAI{[CS
I. Survt:t, of ~ ~ i~ 1.I! 1.~ historical development
. .
1. Introduction •• . • • •
• •
•• • •
. . . . . . . . . . .
2. AtcQistic interpretation of Boyle's law ••••••
3. Atomistic v~ews on the nature of heat • • • • • • •
• • •
• • •
4.. Usa of statistieal concepts in the kinetic theory of gases 5. Statistical intGrpretat10n of the first law of thermodynamics.
6. Statistical 1nterpreta~~oDs of the second law of thermodynamics
7. ~!le HaxwcllBclt7.csnn equation •••••••••••••• 0 .. The BoltzJr.D.l".Jl distribution • • • .0 • • . . • • . . • • ...
9. The Htheorem • • • •
• • •
• • •
. .. .
• •
• •
· .. .
• •
10. The ergodfc theorem • • • .. • •
11. ';i?ne can0nica1. distribution ••
. . . . . . . .. . . . . .
• • • • • • • • • • • • •
CHAPTER II
CLASSICAL STATISTICS
I. g~np.ral J2.t;.Q.P.el:;JE.~" .9.A ~~1ca~ flI.lste~.
1. MOtion as 1so~orphism in phasespace • • • • • • • .. .. • • 33 20 Liouville' g theorem • • • • • ..:.;' • .0 . . . . . . • • •. 34
3. Physically equivalent phases and uniform phase functions • 35
4. Ex:aaple • • .. '. • • • • • • • • • • • • •. • • • • • • • • .36
Page 1 ~ 3 5
7
8 10 lA 18
24 29
• •
Page
• • •• 37
•   40
 .•  ": 41
5 •. :;etrical indecomposab111ty of energy suzrace •
,
6. Poincare1s theorem ••••••••••••••
7. The return time •••••••••••• _ •••
• •
 
1. 2. 3.
4
Existence of the time average • • • • • • .• • • • ••••• 43 Equivalence ot time average and statistical average • • _ • 46 Roptts ergodic theorem ••••••••••••••••••• 48 The ergodic distribution • • • • • • • • • • • • • • • •• 51
. . . . . . . . . . .. ~
5. Systems with weak interactions ••
6. Thermodynamics ot isolated systems
• •
• • •
• •
. . 
• •
CHAPTER III
STATISTICS OF CLOSED SYSTEMS
1. System ot given terperature • • • • • • • • • • • • • • •• ~
2. Canonical distributi·on • • ••••• ••••••• _ •• 63
3. Energy distribution .' •••••••••••••••• _ •• 68
4. Thermodynamics of closed systems ••• _ • • _ _ •• • • .:70
5. Permanence ot canonical distribution in quasistatic
adiabatic transformat1om 73
6. Irreversible transformations • • • • _ • • • • • • • • •• 76
CHAPTER IV STATISTICS OF OPEN SYSTEI~
1, Phase functions and averages for open systems • • • • • • • 82
2. Chemical Potentials • • • • • • • • • • • • • • • • • •  • 84
3. Thermodynamics of open systems ••••••••••••• _ 88 •
4. Characteristic functions involving the t9mperature •••• 91
STATISTICAL T~RMODYN.UlICS
I .
I • SURVEY OF THE MAIN IDEAS IN THEIR HISTORICAL DEWLOPllENT •
•
1. Introduction~ The ai!Il of II statistical thermodynamicsll is to gl.ve an atomistic interpretation of heat.phenomena. In the tlr~t phase of its historical development, the ato~c conception was restricted to the structure of ponderable matte·r· and the atomistic treatuent ot thermal properties of matter accordingly became' known. as. IIstatistical mechanicsll• Later on, however, the phenomenological concepts and laws of thermodynaQics were extended to radiation phenomena, and the interpretation ot these phenoo
. ena in terfJS of electromagnetic fields provided the· basis for a statistical treatment of the thermodynacics 'of radiation on quite the same lines of that of matter~ It seeas therefore advisable to'replace the tradl tional denoeunataon of II statistical mechanics" by a name which indicates more correctly the wider scope of the theory.
In order to cake a first acquaintance vith the ideas and methods of statistical theraodynam1cs, it is instwctive to retrace the history of their birth and growth; we shall accordingly proceed to survey the develop!!lent of the theory, ccnfining our se'tves , . however , to those points which afford the Clost significant illust~ation of the essential features. Thus we leave out ·all discussion of the origin ()f the?· concept of the
.aton in Greece, howeve~ interesting it would be to dwell upon the ancient. atoaists' profound' speculations. Our sta~ting point vill be the revival
•• ~ .. \ ~ ( .~.. ( < •• ~ , "
of atomistic ideas in the tice ot Newton, within the traLlework of the
new lIexperimental philosophy". In fact, the weakness of Greek science
•
2
was its failure to understand the tru~ weaning of thq gucntitativc asp~ct of our acccunt of physical phonomena. !fudern science was born with the recognition of" the paranoWlt iaportance of this aspect; a recognition which could not be expected before the l~vel of technology becaae such as to aake possible accurate experimenting. We shall at , first be concerned ~xcluslvely with the properties of ontarial bodies: the study of radiation is a much later developuent, which we shall also discuss in due cours~.
2. Atomistic interpretations of Boyle's law.
One of the si~plest phencqena to which atomic ideas could be applied was the socalled ITspringll or elasticity of gases, newly discovered by Boyle. The prebleo was to account on atoQistlc lines for
the quantitative relation of proportionality between pressure and den
s1 ty which could be deduce d froQ Boyle I,S mea surenen ts • Newton I s interpretation was based on a representation of the atons as centers 2!
force, a view inspired by his theory of gravlt~tion andwhich remained preponderant until the beginning of the XIXth century. One icagined
that atoas acted upon each other at a~ distance with forces varying
with the distance according to SOLle definite law. The atomic l:.lodel of'
a gas was thus an essentially stat,ic one': the atoms were tl}ought to be kept at definite positions as a result of their mutual interactions
and their interaction with the wall ot the container. The elasticity could be accounted for by Postulating the existence of a centi"'al repulsion between theatoms. Newton* showed by a' very· "staple argument that Boyle's law could be obtained by assuming that the repulsive force between any two atoms varies inversely to their distanc~. ·To see thiS,
let us cl.nsider anuther sample of the same gas, with the saae number
of atees and geoQetrica11y sioilar to the first, the scale of le~gths being nUlt1plied by soaef"actor ~. The total forces exerted on two similar areas of the walls of the vessel will then be in th~ ratio' n~ and sf.nee the areas are in the ratio ~ .~" .the p~e~sures will be in the" .. , '" ratio A,3. rJow, this is just the ratio of the inverse voluaes or den ... " sities, which gives Boile's law. This ingenious argument dispenses US !~~~~~~~~~~~_~~ detailed analysis of the mechanism ot interaction
f Principia (1687)
•
3
ot the gas atoms with the walls or tha vessel. The medel itself, however , involves an arbitrary element, viz. the entirely ad h£.£ aaeuaption of the inverse distance rapulsion between the atoDs.
SaLle tine later, Daniel.Berneulli* noticed that Boyle's law could be acccunted for by a quite different atonic oodel, which was free frOL"1 any. arbij1;rary assunptaon, According to hin, the pressure on the walls of the vessel 1s an essentially kinetic effect: the
. stees of the gas are supposed to ~ove in all directions: when they bounce against the walls of the vessel they tra~fer to them a certain aaoWlt of oOi:lentum.. The total comentum thus received by the unit area
. of the wol1 during the unit of time is the pressurG: althcugh the prvcess of mocentuc transfer is discontinuous, the collisions are ~gined
. .
to take place in so quick succession that the pressure 1s experiencea,
at the oacroscopic scale of observation, as a c~ntinueus force. Nowtonts similarity argunent cay be applied to this oodel as well. The geometrical factors ~ 2,~3 of area and volume are the same, but now the dynacical facto~ ~ 1 is replaced by a kinematical one: in the two sacp1es of the gas, the velocities of the· atoas· are assuced to be the sace, and the amount of nomantum transferred to the wall in each collision is therefore the same, but the numbers of collisions per unit time are in the ratio i\ 1, . sdnce corresponding lengths of path of siallar ·ato~s reaching the wall during this tine are in the ratio
,
A. Boyle's law is thus again established without any analysis cf the
mechanisQ or col11sion ot the atoQS against the wal~s: but Doreover, this kinetic wGchansia, as·alraady.noted, does not necessitate any specific ass~ption 70ncerning the forces between the atoes.
3 •. Atomistic. views on the nature of heat ,
In spite of its greater Simplicity, Bernoulli's kinetic gas
node 1 d1.d not at first· find favour, be ea us e Newton's s tati c model £i tted better into the general scheme of explanation of physical phenoa 1 cna in terllS of central. forces, which was carried to,. its utmost ccnsequences during the XVIllth century. A striking illustration of the
way in which i1central torce atomismu was developed 1s offered by the
.
* Hydrodynamica (1738)
4
interpretation of the heat phenoQGna, whose qunntitativ~ anolysis is the work or XVIllth c~ntury physicists, torcQcst oacng them Joseph Black. _Heat was thought of "as an inponderable fluid, the "coloricll, Whose a tC!'lS cC'uld ca bound by specific centl"ol forces. to those at ponderable matter. Names still in usc3', like IIquantity o~ heatH, IIheat capacity" owe the~. origin to this sUbstantialistic ccnoeptio~.
y~t here also therG WQS a l"iva~ view of the nature vi heat, based on a kinetic picture of D3tter. H€ot could be conceived as the result, on cur macroscopic scale of observation, of the individual p,otions or the atoms; if one takes the tGnperature as a Ileasure of the average velocities at. atcaic nct1ons, transfer c:t heat frctl a hot to
,
a cold body would be' a ~ansfcr of k1nct1c energy resulting fro!] the
.interactions cf the atoos ot the two bodie·s. This theory at heat transfer is just as Simple and accurate as the substant1alistic una, acccrdi~g to which the transfer of heat 1s a flow of caloric frc~ a higher to a lower "1evel" .(indicated by the tcr:aperaturc). In :rect, as was pcinted out by .Lavoisier and Laplace* the two theories, so far as heat transfer goes, are perfectly equiv~lent.
The superiority cf the kinetic theory of heat became apparent, however, in another'f1eld, the exploration of which was initiated,
. 
du:ting the first.' half ot the XIXth century,. by the introduction of
stean as a s0urce of' power in industry. This focussed the attention on the transf'ornations ~:r heat into ~echanical work and ~ versa, and eventually led to the :ror~ulction of the law cr c0Dservation of energy as a universe! principle, gove~ning all transforcations of phySical agencies. Rumford was the first :tc pcint cut hew difficult it is for the .ca1oric conception to account fer the pr.9ducticn ef
heat by friction, while, the kinetic theory yields an obvious interpreta tion of this phenonenon. But it was above all Joule** who followad up the consequences of the kinetic thecry cf heat in this respect and established their truth by rigorous experimental tests: he shewed
 how the law of conserve tiori of oechanical energy, applied :to raIl the
"
* MEh.1oire sur 12. chaleur (1780).
** Joule's great work was carried out in the period 184043
5
Iilot;1ons of' the systcn, including the intrinsic ateme t1o~lGns llsnifcsted as heat, could be expressed as a law of equivalence betw~en quantity of haat and mechanicolwork.
This ncocntcus disceve:.:y was a striking dCl1onstration of the heuristic powor of the kinG~ic theory of stcuic ccnstitution, and it established the vaLlue of this theory as a principle of' unifiea tien (·f
cur world picture: the laws gove:.:"ni~lg the phencnena of haat were now , "reduced" to these of cechanics. At this stagG, Q word of \lorning will not be superfluous: it would ~ wrong to conclude (as Jcula did) that the experiC€ntal verification of the law of equivalence between heat
and work prcves the truth of thekinetic theory. As a catter of fact, the law of equivalence i~ ~ntirely independent of any particular view
on the nature of hGat; this point was especially stressed by Robert Mayer*, who at the sa~ time as Joule arrived at the saae ccnclus~on
by an argumGnt.just.as rigorous as Joule's and exclusively based on e~p1rical relations fornulated in teres of oacro~copic quantities (such as pres sure and s pe:cif'ic hGot).
Mayer always objected  quite rightly  to the phrase "heat is Q node of' motionll in which cne often scught to ccndense the kinetic ccncepticn. Indeed, heat i8 and remains a physical phencncncn quali
" tct1valy different fr(.;:} that of Llotioll in bulk, and the establisbnent of a _quantitative equivalence between the two phenoDena in no way affects thiS quali~atlve difference. The law cf equivalence suppliGS
us with a quantitative estincte of transforQat10ns of quality in physical systccs. "Joule and Mayer never understood each other; each th<..ught that the othc;r wa: hopelessly wrong. In fact,they were both right: they €cphasized different and equally ioportant aspects of
the equivalence between heat and work.
4. Use of statistical concepts in 'the kinetic theory cf gns~s.
Let us now coueback to the kinetic picture of an ideal gas "
as a  sy.steo. cf a teas IilCving freely (axc_opt for .cccasacnaf co111s10ns),..;_,·" inside a vessel, against the walls of which they are reflected. We
* Mayer's first paper was published in 1842. "
have seen th!lt B(;yle's law is nn ir:l!lGdiate property of this model, but we are now in a position to enquire about its further consequences',
I
especially concerning the thermal properties of the gas. This invest~
ill I t"
, gatlon was, historically, initiated .by Joule himsl?l.f and pursued wi~h .... "
greater mathematical skill by Clausi~s and Maxwell. ~he first pro :.
perty to be considered is, of course, that expressed by Charles' law, 1 .• e. the proportional! ty of the (invariant) product of pressure and
_volume with the absolute temperature of the gas. A more detailed analysis of the atomic mechanism by which the pressure is produced
. .
will thus give us a precise kinetic interpretation of the concept of,
(absolute) temperature.
An atom of mass . .!!!, on being elastically retlected by the wall "".: normal component 01' velocity ~, yields to the ~all a momentum
Let N be the total number of atoms contained in the volume V ,'.'
 
6
with a
2mv.
..
and f(v)dv the_ traction of these atoms with a normal velocity component in the interval (v,v+dv); the numberot collisions of the type considered on the unit of area of the wall perunit time will be' (N/V)v f(v)dv. The pressure f will thus be obtained by integrating the expression Z(N~/)cv2t(V)dV ,over all values or ~ or one Sign (ind1 ~ating an initial motion towards the wall), or halt the expression
,
just written over all values »s z Therefore, I ,'
PV' = Nm <../) ,
where the notation'
(,/) =J t(v)vZdV

representing the average value at the function";' tor the distribution '.
ot y given by,f(v)dv. If K is the kinetic energy of. an atom, we have
obviously
,
and therefore
= ~N(K>. 3
Comparing with Charle~· law, we see that the absolute temperature is a measure of the average kinetic energy of the gas. Besides yielding
p V
7
this kinetic inte~pretation of temperature, the above a~alysis illustrates the general fact that the definition of macroscopic quantities in terms of corresponding atomic quantities essentially involves a statistical average with respect· to a distribution which characterize~ the state of the atomic system. This statistical aspect of the atomia
, I
theory of matter *as first clearly r~cognized by Maxwell, but it is '
inherent in the very conception of an interpretation of observable phenomena in terms of entities not accessible to direct o~servation.
·ObviouslY, such an interpretation would be meaning1ess if it required an accurate lmowledge of the dynamical behaviour of the individual
atoms.
5. Statistical interpretation of the first law of thermody
namics.
The example just treated is an extremely simple one, and
the distribution function occurring· in it is of a very special type.
It shows us, however, how to proceed in the most general case. Let
us consider any system of bodies in any physical state: _from the atomic point of vie,'!, we have a dynauicolsystemot an enormous humber of degrees of freedom, whose state is described by canonical 'co
ordinates ql' ql'···' qr; PI' P2.' ••• Pf' and whose behaviour 1s determined by the Hamiltonian function of these coordinates. The Hamiltonian will further depend on certain macroscopic parameters 81,82, ••• which define .the extarnal conditions to \oThich the system is su1:jected;the values and variations of these paramEters are under the observer's control, in contrast' to those of the· atomic coordin~tes q,p. We accordingly write the Hamiltonian as a function E(q,p;a) and it will
of course suffice to consider a single external parameter ~ (e.g. the volume y. of the vessel in the case of a gas).
While the Hamiltonian determines the state, or phase, of the system on the atomic scale, its macroscopic state will be characterizc:d by a distribution density t(q,p;a,O)dw, which gives the pro . bability to f~nd the system in the elei1l€nt dw: dq , ••• dqr dPl •• dp'f  of phase space aroun~ the phase (ql •• qf; Pl ••• Pf). The distribution t depends' also on the external paraaeters al, al, ••• and moreover
8
on a parameter g (e .g. the temperatureJ which charac'te it zee the state in question from the thermal point of view. The macroscopic quantity which 1s called in ther~odynamics the internal energy U of the system is then expressed as
U =. (E) = J E • t dw •
(1)
According to the first law of thermodynamics, any va~iation dU of
the l~ternal ene~gy can be refer~ed either to external mechanical work S:: per.formed El the system by varying the parameters /!, orto a quanti ty of heat,) Q, supplied to the system by varying either a or g 0;"
both; i.e. we have
•
dU = 3~ + J'Q,
(~)
and Se 1s an expression of the form! da, a being some function of a,Q. If, for example, ~ is the volu~e ! of a gas, A represents the extermsl pressure f which balcnces that of the gas; quite generally, A 1s called the (generalized) force associated with the parameter i. •
. Now the Hamiltonian, as a function of a, is so defined that the work ;:. ... E
done t:y the system ~lhen .9. is varied is expressed by  as da;.
therefore, we have immediately
A e (.£Lj J ,,= I ~E 'da ,
.a· , '. as/'
as the kinetic interpretation afthe concepts of f'o~ce and work.
Froe this, the interpretation of the quantity of heat a Q follows readily.. In fact, we get from (1)
dU ;: J dE·' fd W + [ E • <if • d W ; the right is just &~, ther~fore by (2) the
SQ = J E • elf • dw.
the first term on is d'.Q:
second
(3)
This formula is important in that it shows that ·the concept of heat is very directly connected "Ii: th the change in the stot1stical distribution due to the change of e~~terna1 conditions •.
6. Statistical interpretations of the second law of thermo~ dynamics and the problem of irreversibility.
The second law of thermodynamics has two aspects. In the first place, it states the existence of a thermal quantity, the
•
::> ... '
9
entropy, which (in contrast to the qu~ntity of heat) is, uniquely defined (except for an arbitrary additive constant) for each stote'o! thermal equilibriumof a physical system.~ The entropy difference
dS between two neighbouring states ot equilibrium is defined as
dS = li.., (4)
I . T
aQ being the quantity ofheat which must be supplied to the system •
in the initial state at t~mperature T, in order to bring it into the tinal state in a quasistatic way, i.e., by controlled variations of
the macroscopic variables •. The second law then makes the further statement that any spontaneous transformation ot a thermally isolated system produces an increase of its entropy, indicating its·itreversible character.
The comparison of formulae (U)·and (3) shows thstthe kinetic.interpretation ot entropy will involve statistical conceptions
in a much deeper seDse'than is implied by just taking statistical overages: a. change of entropy is. directly. connected with a change of the stDtisticaldistribution ltself~ The fundamentally statistical meaning of the entropy concept 1s still more apparent in the case
of irreversible proces~es •. At first sight, we meet here with a' parada •.
:leal. .. si t_uatton: .ha\,1 can .an. irrever.sible. . evolution· be. ·dei"ived.·froa ·a.pureJ dynaQ1col model, whose behaviour is essentiall~ reversible in time? This is the central proble:m of statistical thermodynamics and its general solution is ~y no m~ons easy •. It isclear, however, that'it can only be found in the properlY.stotlstical feature of the atomic picture; in fact, it will turn out th~t thermodynamic irreversibility. is an example of a very general property of stochastic processes, viz. the tendency of such processes to "tend" (in a sense peculiar to statistical_theory) towards the situation of greatest prebability.
. A simple exomple will illustrate what is mennt by this.· Con·
•
sider the spontaneous mixing of two gases:
the increJse of entropy
in this process,'when considered from: the atowistic point of view, can 'Only appear as a ',function' .ct the numbers of atoms involved; and it is in fact Simply related to the gre~ter probability of the dis
10
ordered stote .rzpresented by the mixture, as compared with that of the separated gases. More generally, heat in its various manifestations is the r€.sult of [1 disordered form of motion of the atoms (in contrast to the ordered bulk motions on the macroscopic scale); entropy is 3 st~tisticcl ~easura ot this disorder, ond its increase expresses a tronsit*on to a state. p£ greater probability.
.
7. The HaA'"Well.Bolt zrJann equation:
The elucidation of the statistical aspect of the second law ot thermodynamics was the joint \'lork ot Maxwell and Boltzmann; the contributions of both great p~ys1c1sts to this dift~cult investigat10n are equally essentia~, but Boltzmann was perhaps more actively engaged than Maxwell'in' developing the most general implications of the statistical point of view and in detending it against various objections. The first problea to be tackled was the simpler .one ot the ideal gas;· afterwards, the case .of general dynamical systems ot atoms was treated end required for its solution the elaboration of
new ne thcde ot . approach. We shall now dwell. at; soae .length upon the r. discussion of the ideal gas according to the original method of l~x
" ~
well and Boltzmann; in spite of its limitations, . this method is still ~
• F
of considerable 1nteres.t and l::rings out the fundamental p01n~s very
clearly.
Let us consider 'an ideo1 gas, i.e. an assembly of atoms assumed to be without interoction except during. the short1ntervals during which they collide in the course ot .their random motions. The simplicity of this case lies in th~ c1rcumstance that we caD treat each atom separately and define a distribution function fOl" its pos
,
" .
sible dynamical states indepondently of the other atous , In other .....
words, ''Ie may in this special case apply sto. tis tical conside:cations to the individual atoas and not only to the toto1 sys.tcrn. In order to cover the case of iireversible transformations, we extend the definition or the distribution function to nonoqui1ibrium states and we according1yas$ume that'the distribution dens1 ty t(~ .'' 't)dw depends on the tif'Je e~plicit1Y (and not'only t~ough the position vector"t' and !,lomentuq p of the atan?; here dW 1.8 1;be elecent of
,'.
 + .'
...
t , .
~. ~
r ••
..

i '
11

phase space of a single atom, i,.e,. the product of the volume element dv of ordinary space and the volume element dvp bt momentum space,
,..." , dW:: dv dVp. It 1s convenient to normalize the distribution density in such a \"lay that ./fdW= N, the total number of at cms ,
As alr ':aq.:y stated, we neglect the interactions between the ., atoms (for their ruut.ial colliSions a schematic treatment will prove
~.
sufficient); we,.may, however, withOut g~eat complication take account
or any external torce (such as, e.g., gravitation) acting u,oI1 each
~_ atom separately and upon all of them alike. We shall assume that this t: force is def'ined by a potential vf.:r) and we may thus write the Hamil
~;,' tonian of' an atom intbe form
~, . H '= K. + V, K = im pi! ,
where, for simplicity, pi!. stands for IJr I~. The equations 'Of motion (except for the effect of col11sions) are then
dV
~  grad' v.
dt
'.
Our a1m must now be to study the way in which the distribution fWlction varies in time, as a rEl!sult at the collisions, and to show that it does IIten~Utowards a de:1"inite form which is stationaJ;"Z, 1.e. for which af/ at = 0 for every phase. The, definition of the entropy will suggest itself in the course of the argument. Let us co~
~~ sider the fdYJ atoms originally occupying the element d~ of phase space; after a time interval dt, it ~o c~11i$10n had ocurred, they would be found around the phase ~'I: x'>+ (p>"/m)dt, p '= p  grad Vdt within a region whose volume dUlf, to the first order 1n dt, is equal
to d,.)J (as 1s eas1ly seen by forming the Jacobian ot the variable transformation). Thus the difference
.."
d.f'  f(~ + £ dt, P  grad V. dt, t + dt) f{X'; p, t)
would vanish. However, as a result of" the collisions, a certain number Qdw dt of" atoms will enter into_the element dtu and a numbe;t",: bdUJdt will be knocked out of" it. ~heref'ore,
·df
 = ab.
dt
(~)
This is the fundamental equation ot riaxwell and Boltzmann, which governs the variation,ot the distribution function with time. The lefthand
side can be written more explicitly
df ~ f .
dt =. ;r.+ grad t • ;  gradp r •. gr~dV, (6).
where gradp f ·denot~s the vector with components If~l1: .•. The main task 1s to estimate the difference ab on the right.
At this stage, statistical considerations must be introduced.
We cannot hope to evaluate ab exactly, since to do so we would requ1r~ an exact solution of the equations. of ~ot1on, including the effect 'of the collisions. But we are not s·ctually interested in such an exact' calculation; all we want to ascertain 1s the average trend of the distribution.  For this purpose, it 1s sufficient to replace the difference ab at time .t' by its average value. The latter is readily expressed in terms of the probability per un1ttime for the occurrence
of a collision of a definite type: this probability is the only quantity representative of the collision processes which will enter into
·our argument; we shall not requ1re any explicit expression tor it, but only use a very goneral sym£.1etry property whichi t mus.t have.
In the first place, we must characterize the d1rrerent, types
of elastic col11sions. We need not make any definite aSbumption concerning the law of interaction of the atoms during a collision; we assign as space coordinates to the colliding atoms those of their center
·of mass, and any further description of the. mode .. of" coll1sion 1s contained in the abovementioned probabil1typer unit time. The col11sion 1s then entirely characterized by the values of the momenta of the atoms .
. ~ . ... ~
before and atte~ their interaction; these quantities PI' PZ; Pi, P2
are only restricted by the conservation· laws
1i + ~ = Pl1 ~ pi Z ' PI + P2 = Pl c. + p~c!
(7)
..
In the barycentric system of reference, the momenta of the two atoms before and after the collision· . are' equal and opposite, and their absolute value 1s not alter~d by, the collision; the only effect of the latter is to alter the oriantation of the common direction of the momentum vectors with respect to the plane defined by the ·initial mo
13
.. adW
menta in the "rest" system (in which the gas is at rest)., To deter:aine the type of collision, we may accordingly choose as our 8 independent paramete~s the two angles which define this orientation and the two momenta in the initial state.
The probability per unit time for this typeot col11sion will thus be of the J t orlll
li(~Pc ._'" pi 12. ) dv dVPJ. dVpt! d ~I
where d $1 is the element of solid angle for the direction of the final momenta in the barycentric s~ystem. We shall also have to consider the "inversell collision, in which the initial and final states are interchanged; the corresponding probability per unit time is
w(Pt P2"~ Pa) dv dV~ dVpz d~ •
It is readily veritied that, in virtue of the relations (7),the elements ot variation of the po rllmcters are the s:l:me: dv P~ dv P d~ 1 = dv p' dv P 1 d~ I • \'le ass~, moreover, that the probability densit9L~ ~s also ~e ~me value for the inverse col11sions; this assumption of detailed balance is only fulfilled for the simplest kinds of collision, but Since we
are here only concerned \>li th the essential features of the theory, we
i
may disregard the complications which arise when it has to be replaced
by a less restrictive assumption.
We have now, for the average values of a and b,
(:tt~1 ................ ) (.+ >, .) (,. '>, t) d ih
= ;w P1P2·P1 Pz f x, Pi,t t x,PZ dV dVpiVPad:t,
/
(
= \ (>...... ~ ..) fC ...... '> t) f( ...... I" t)d d d d~'
)' w P1,Pz "'Pl Pc x, Pi'  x, Pc' v vp . vp ~,
12.
bdW
where we must take Pi  p, and the in.tegrations must be performed ~
over the momentum parameters so as to l~ave Pl confined to the volume
element dVp • These formulae are often raterred to by the name Boltz
, 1
mann gave them: the nStosszahlansatz" ,i.e. the assumption on the
(average) numbers of co~l1sions. For h we get immediately in somewhat
Simplified notation, .
b f (4 " ...., ~ ) 11'(~) f() d.1\1 En: =_ 4pJ(8)
= W P} P2 :... PiP 2 .. P1, Pz dv P t! 'i!, .. J.
11
14
where the integration extends over the whole range ot momentum p. and
. Go
angular coordinates pertaining to the solid anBle p1. In order" to'
ohtain the corresponding expression for ~,we have only to make use .' of" the detailed balance assumption, which gives
C'. a J (  '>1 91) #(~) #(+1) d d.i\, ":i' = . 1 (9)
: W PIPZ·fj. Pz .L Pi ~ Pz vPz J: • LVI  P J
.  :;.;.. _.
8. The Boltzmann distribution •.
It in the Haxwell ... l301tzniann equation (~) !. and .k are understood to be the averages just :calcuiated, this equation is no longer an exact dynamical re1at1otl, but oan only claim to represent the average trend ot the distribution function: this, however, 1s just what we are interested in. We shall use it firstto derive an expr€ssion for a stationary d1$tr1but1on, and then show that no~her distribution can be stationary; this second step will at the same time provide us with a complete kinetic interpretation of the second law.
Sufficient conditions of stationarity are obtained in equating to' zero the two sides of equation (5) separately (and putting ~t/ at = 0). Further, the condit!onab=o will, be satisfied, according to (8) and (9), it we impose for any po1nt ~ of space the more . stringent condition
f(Pi) t(~) = t(Pi) f(p~) (10)
tor any set of momenta satisfying the conservation equation~ (7). This will determine the way in which t(St,p) depends on' tor any ~siue
of X. • Then, the cond! tion cit/dt :: 0 with atl"t = 0, 'i.e. by (6)
.....
grad!. E  grsdpt • g~ad V= 0 (11)
m ,
will_yield the dependence of f on:t.
Let us apply the condition (10) in the carycentric system,
'. ...... ~  ~ ...
in whl.ch Pl =,,p~, pi =. P2' the absolute values of PI and pi or,
if we~ike,' the squares pf, pti , be!?g equal,.. This shows us that
the product f (V) t ( _1>1
is inde pendent of the ciirection of"t and 1s a ~unct1o~ ot pl. only. But in this system, we have obviously mirror sy;ru;.etry in momentum
. '
,: .. ;.. ~. ,
15
space, i.e. f(p) = rep). \tIe may therefore write fep) = g(pz, • Now, let us write the condition (1 ) for a particular collision in which Pi ; 0, and therefore P~ = Pi2 + PZ2• Putting tor a moment
~". p!2 =::. , pz.2 = "(I ,this gives
~.
I '
r (j) g ('t'J) = g (0) g.(}+1}).
taking the logarithmiC
l
'''' This f'uncional equation is easily solved by r derivative '"11th respect to .,. :
""'
g ,I, (r ~_ :: g I (, +.,~ ,
g(~) g('+~)
which shows that this logarithmic derivative is independent df p' ahd ••. , reduce s to a f'unction of'.~. Therefore', t <it) 1s of' the form
L_
7: . t6t;'P) = 1'0 (X') eXPl  ~'x) K ),
where the variable pc has been replaced by the kinetic energy K. 'We must now insert this expression for t(x,~) into condition' (11), which gives
f · [ + grad to + ~ grad V  K grad P ) = 0
. r , .
L r'
Since this equation must hold for all values ot ~ we have separately
grad II = 0 and Fo grad to + II grad'V :: o,
Thus, ~ is a constant and fo(;t)c const. exp (~V). Collecting all ~ ...
the r~sults, we obtain a stationary distribution density ot the
form ~~
. t<1';P') = 4 e~( ,p;a)
(lL)
where E is the Hamiltonian or the atom, which also contains the external parameters ~, and where we have put
Z = J eJiE dw (13)
in order to norm£lize t~e distribution. Formula (12) is called the Boltzmann distribution; it 1s a generalization of that first derived
•
by lfaxwell for the case of an ideal gas with no external forces.
Assuming ~rovisionally that Boltzmann's formula represents .. the unique equilibrium distribution of the ideal gas, we must still ascertain the physical m~aning of the parameter~. This will result
16
from a s1mpl~ r~lotion between ~ and the average k1natic,energy (wh1c~as we know, is proportional to the absolute temporatur~ For any
~ . component Px ot the momentum 'it; we have obviously
"" 2 {pi exp( /3 ~l>~)dP "PO Pi
(.:]!L~ N .. ~ 2m J 2m x = N ...iL log I e~~ dp ::;
~, z.ml r" Pxl iP ~ x
~< J_ cxp(~ Tm) dpx
~ ell
1I$i.:..r.:..:" ..
H·:t~
or finally
'/ t: 2 .
*" log (~»)., ax dx =to N ~ log (i"".)
00
2 .
1 < Px \. _ 1 A 1 _ 1 * T 2m/T .... Te'
(14)
;fir. = .... N
if we denote by e the inverse of p. Therefore,
. (K) = 1 N e • (15)
Now, remcmber~ng the relation, derived above, PV::; ~ <K) ,and combining it with Charles' law PV = RT (where R is the "gas constantlf),
~ ..
we get
(K) = 1 RT •
The comparison with (15) gives
e = kT,
(16)
with k = R/N. We thus see that e can be regar~d as a dynamical measure ot the absolute temperature.
However, for such an interpretation to be at all acceptable, it must be universAl, i.e. independent of the nature of the
gas (which is represented by the atoillic mass m). Now, the universal character of the fundamental formula (14) may be verified vy extending the preceding argument to a mixture ot two ideal gases: on account
of the additional condition ot the form (10) for collisions between two mol€c~es ot ditferent kinds, it will be found that the Boltzaann distributions for the two gases contain the ~ parameter~. This remarkable property is described as thG eguipartition ot energy. We shall presently discuss its wider implications, but tor the moment
•
17
we are only concerned with its bearing on the ideal gas laws. In
t~e first place, it illUlsdiately provides the kinetic lnt'erpretat1on
of the universality at both the absolute temperature (embodied in
the empirical fact that all gases have the same thermal dilation coefficient) and the gas constant. With regard to the l~tter, we con~ elude that it has the snme value when referred to the same number of molecules of diftej,:'ent gases. Hence, the coefficient ~ is (as it
should be for the consistency of our interpretation of e) a universal constant, called Boltzmann's constant. The property Of R just formulated corresponds to th~ empirical law, discovered by GayLusaae, that R has the same value for all gases when referred to a mass of one
mole: we infer from the comparison of the two statements that one
mole of any gas contains a definite number A of molecules. This property ha4 been postulated by Avogadro to account for the laws of chemical combf.nc ti on of gas se s; the kine ti o law of e quiparti t1'on exhi bi ts
its close relationship to GayLussac's law. The value of the universal constant A, ca~ed Avogadro's number, can be ascertained in an indirect way by a more detailed kinetic analysis of the nonequilibrium, properties ,of gasses and the deYiatlons from the ideal gaseous state*; ; we thenget the numerical value otBoltzmannls k as R/A, where R now denotes the universal value of the gas constant referred to ahe mole.
We quote the modern figures:
R = 8.314 joules • deg1• molel,.
A = 6.023 x l~3 moleI,
k = 1.38 x 1016 erg.degl •
.( 11 t .
.'
 ..  ~ ..... ~ ~ ~ ...  .~  .....
* The study of nonequilibrium states (eg.diffusion) yields the mean free path, or average length.of path between collisions: this 1s inversely proportional to the pzoduct. of the number H of molecules pe r. unit volume and the square 0'" of the i'diaaeterll of a molecule. On the other hand, the total volumeNN ~3 of the molecules was first estima
ted by Loschmidt from the volume of the liquid into which a unit vol
ume of the gas condenses; a better estimate can be derived from the volume correction of van der Waalsl equation of state for nonidealgase~ ,
,
18
" The Htheorem.
'.~.
Boltzmann's proof of the uniqueness of the stationary dlstrir~ but10n (12), 1._e. essentially that the conditions (10) are necessary
"~' tor the equilibrium, is one of the finest jewels ot theoretical physics. ; The idea 1s the folJ;owing: consider a functional H[f] ot the distr~
=_ butioD; its variation in time as a result of the collision will be
 some functional of both ~ and df/dt,
.9!L = G ! 1" 2f J"'
dt 'dt'
".,

in which df/dt must be replaced by~its expression a~b. If we use the
, average values (8) and (9) .of St and s., we shall obtain the average trend of the functional H in the form
~ = F [f(Pi) ; f( Ii) f(pZ) f(PiI) t (~) ,
where F again denotes the appropriate functional. If now we can construct H in such a way that F never vanishes unless the conditions (10) are fu1r111ed, we have obviously achieved the desired uniqueness proof. Such a functional H serves, so to speak, as an indicator of the deviations of the gas from equilibrium. On physical gr.ounds, we would expect that once H \lould have r.;:ached its extreme value, it would remain constant, corresponding to the fact that when the gas has reached its stationary state, it remains in this state indefinitely. In other _words, we expect the variation of H in tiue to be monotonic until a constant value corresponding to the equilibrium is attained.
Boltzmann has shown tkat the functlonal,~hich represents essentially the average value of the logarithm of the distribution density exhibits the expected behaviour; this is the content of his famous Htheorem. In precise terms, let us take
H =J flog (f·oUl) de..>;.
(18)
1n this formula, tAl represe~ts a "unit cell" Of_phase space, of arbi
19
trary volume. It has been introduced 1n order that the quantity
.: whose logarithm is considered be dimensionless, ahd 1 ts inclusion means that H is only defined up to an arbitrary constant N~, which is ot course of no influence on its time variation. We have then for the average time variation of H on account of the collisions
I . dH
Cit ~' 0 , (19)
the equality holding if and only if the stationarity conditions (10) are satisfied.
The proof makes an essential use of the assumption of detailed balance. Taking account of J (d£/dt) d,...,= 0 we may write
dH fd£ I
tit ='j Cit log t d(...' = (a  b) log t d&rJ
=1.[ wCPiPZ"P'l pi) (fCPi') tCP'z)  fCPi) tCpZ)} log r(Pi»)(. dv dVP1 dVPZ d~t •
Since Pl and 1~ play· entirely similar
dH 1 r· )
<it = 2. J W • (fnJ fp'  fp fp
. .~ 2 1 ~
p~rts,
log (fPlfP2) dv dVPl dVPZ d~' •
Interchanging the initial and final st~tes, and taking account; of the detailed balance assumption, we finally get
dH 1 ~ [ 1
, dt = 4Jw. (fPi fp!  fplfpZ) log (fP1fpZ) log (fptfpZ) Jdvavp]. dVPa d~'
,The truth of the statement (19) is now apparent, owing to the monotonic character of the logarithmic function.
The H  theorem, however, has a much deeper significance than that ot a mere formal instrument to establish the uniqueness ot the equilibrium distribution. ~he property of the H functional which it expresses immediately invites comparison with the property of the entropy expressed by the second law of thermodynamics and suggests the
..
existence of a close relationship between the H functional and the
entropy. We shall first show that the quantity H, taken for the stationary distributi9n, may be ident1t1ed ~nth the thermodynamical entropy of the ~ystem in the corresponding stateo! equilibrium. Strictly
•
, .
, 
20
speaking, this stationary value Hstat' which is a dimensionless quantity, should be multiplied by the conversion factor ~ to give the en~.tropy in the usual thermodynamic units; but it is more convenient to
.. use the dynamic measure 6 of the temperature and accordi,ngly treat the ~,. entropy as a dimens:i{onless quantity •
. : To see this, let us write down the expression tor Hstnt~
.. ""
 Hstat = N log (Z/l'fw) + ~ (E)
(~O)
· and compare it with the thermodynamical formula
e s s U  F,
(a.)
.t where U ii (E) .and Fdenotes the tree energy. We see that the identiti
,
cation at Hstat with S is equivalent with the identification
F =  N e log (Z/Nw) ,
;;' or
which, il1cldentall.y, gives a siL~ple physical meaning to the nor.i1lalizatien ip.tegral Z. Now., to justify equation (a), we have to show that the function F defined by it plays the part of a characteristic fUOc, .ll.Qn for the independent variables 9, a, i.e. that for quasistatic
transformations dF =  S de + Ada,
.ez, = S £L = <.2L.> .. ~
,99 ' da .aa
· The last formula is an immediate consequence of (zz), As to the first,
'+
'we may eliminate S from it by means of (.21) and transform it into the
Helmholtz formula
0" = .iL (i3F) • .;tp
· This last equation' is r~ad~ly verified to r,z;su1t from (a), which com
e pletes the proof of our kinetic interpretation of thermodynamical quantities.
Passing finally to spontaneous, irreversible processes, we see  at once that the Htheorem implies the law of increase ot the entropy
.. of an isolated syst~m. ' The fact that the Htheorem gives the average trend of the quantity H shows that the increase of the entropy is an
,
essentially statistical property, and the irreversibility to which it
corresponds has a statistical origin. The successive values actually
= ..... 1IiI'"
21
taken by H in the course ot time fluctuate aro~~d the averages following the trend of the Htheorem and, strictly speaking, this behaviour
is r'eversible: evan when the system has reached its st£ltlonary state, the fluctuations of H around its constant average value continue 1n. definitely, and we cannot exclude the spontaneous reappearance of any i state, however far/trom the equilibrium. However, for systems of a large number of degrees at freedom, large fluctuations will be so 8e1 dom.ns to be pract1cally impossible: for such systems, statistical
. ,predictions such as that expressed by the second law of thermodynamics become practical certainties. Fo,: smaller systems, on the other hand, as for instance emulsions of submicroscopic droplets, deviations from the thermodynamical ,behaviour do occur and the fluctuations eventually become so preponderant as to exclude entirely the use ot thermodynamic concepts.
In this connexion, it is important to realize that the kinetic viewpoint is much wider than that of thermodynamics. The functional H is defin~ £or every state o£ the gas, whether .in equilibrium or
not, whereas the entropy is only defined for equilibrium states. This nnturally lends us to enquire whether the quantity H has any physicn1 meaning also for nonequilibrium states. The answer given by; Boltzmann to this question is extremely interesting; in particular, it a£rords
a still more precise elucidation of the statistical character of entropy. Let us consider any distribution o£ the atoms rrom the point
of view· of it~ prop~~~1J~Y of occurrence. For the evaluation of this probability, we may assume that the ~ priori probability of any "cell" of finite volume of phase space is proportional to its volume; we
: have seen, in fact, that this volume remains Wlaltered when each or
its pOints moves along the corresponding trajectory, even if the effect o£ the collisions is taken into account, Let us, therefore, divide
the domain ~~ of phase space, necessarily finite, within which any atom remains in the course of time, into arbitrary cells of finite volumes LJi; a distribution 1s determined by the numbers Ni of atoms in ,these cells, and its probability is given by
P (Ni> = NJ .!J (~,lii (24. )
rt '1 {~ 'I ,
'i Nil J. ~
~ I
.1
G2
where N = 2: N is the total number of atoms. If we assume that all
i i
the Nils are sufficiently large integers (which obviously excludes only a few very unlikely distributions), we way write approximately (using Stirling's asymptotic formula)
~ "Nw
1l0g.p (N1 > ~ N log IT  ~ Ni log ~
•
In this formula, the sign"~ 1s used to denote the asymptotic character
; ,of the approximation; the arbitrary "Wiit cell" w 15 again introduced for the sake of dimensional homogeneity. Now, if for the purpose of mathematical computation, we go over to infinitesimal cells ~i~d~), the ratios N1/~\1 will reduce" to the corresponding values of the d1s
~ tri bution density .t and the term L N i log (Ni w IU].) in the last
" 1
formula will just reduce to the functional H. We see, therefore, that
apart from an arbitrary additive constant, the quantity H if] measures the logarithm of the probability for the occurrence of the distribution t in the course of time. If the distribution is a stationary one, the lc~A"'1.~hm vf its probability of occurrence represents the entropy of the corresponding equilibrium state.
From our present point of view,however, the equilibrium distribution, instead of being characterized by its stationarity, can be defined regardless of temporal succession in a purely statistical way, cviz. as the pistribution of maximum. probability for the isolated
;
" system 'mdcr consideration. To show this, we must f'ind the extremes
of'the functional H~] tor all distributions satisfy1ng the conditions
J fdw = N and J ~ t dw = U , (25)
which express that the gas 1s an isolated system. By the usual method ot Lagrange multiplicators, we find
Jdf · [log t + ~ E log,,]dW= 0,
t = o(.e13E ,
,,.. 
..... ~o:::.,
where the para~Gt~~s~,p are determined in terms ot N and U by the
" accessory conditions· (2'5>. These give '\
0(. = N/Z with Z = r e~dW
23
and ~.: (kT)~l as before. It may be observed here that in all formu " lae involving the equilibrium distribution we have tacitly extended
the integrat10n over the mOlJ),enta to the whole momentum space, although·
(as pointed out in the preceding argument) the phases actually occupied  by an atom in the course of time are necessarily restricted to a finite 'dom~ln .r" ... of phasJ s.pace: owing to the exponential dependence of the ::..d1strlbut1on densi ty on thi.? kinetic energy, this procedure only 1n
: volves·a quite negligible error. More~ver, it 1s clear from either
 deriva~lon that the exponential form of the distribution density is, , just anasymptotic one, valid only for systems of a large number of
atoms.
vIe have thus established the asymptotic equivalence between the stationary distribution reached by the system in its temporal evolution "and a purely statistical distribu~~on over the available domain of phas ·space, fro~'whlch Gvcry consideration of temporal succession has disappeared. In other words, instead of taking averages over the ,phases actually occupied 1'n succession by the system, we take them ,over all the possible phases, each being weighted' accordlng,:to its probability of occurrence in the ten:poral evolutton. The equivalence of the two points of view becomes of f'illldamental importance for· . ,the. extension of statistical thermodynamics t·o more, general systems of at oms ;. such as chrystals, in which the interactions between the atoms play an essential part. It is then no longer possible to study in detail, as we
did for gases, the effect of the interactions on the statistical dis'tribution of the dynamical"states of the system, but itturns out that the equilibrium distribution can be found,just by following .up theidea~ of the equivalence between time· averages and statistical averages. This is in fact the very foundation of statistical thermodynamics in its
most general form, and we shall have to discuss it thoroughly later on •
. For the 30ment, we shall ouly outline the, arguoent in the still imper
•
_ fect, though substantiallytrue, form in which it was originally de
veloped by'Boltzmanri and .... Uaxwell ..
24
10. The ergodic theorem.
Since the Hamiltonian of the general system we are nO\1 con . side ring is no longer separable into a sun of terms pertaining to the individual .atoms, statistical considerations must not no~ be applied to the asseobly o~ .atoms~ but to a (fictitious) asseably of IIcopies" of our total system, all placed in the same external circumstances.
A llhase or dynamical state ·of the system will be reprasented by a pOint in a phase space of 2f dimGnsion~, if l is the number of degrees of freedol!l of the total system. A st,atistical distribution in this phase space will refer to our fictlti~s assembly of systems, i.c., it will indicate the probability of finding anyone systec in a given "ccII"
of phase space under the·given external conditions. In the course of time, the rcpresenta~lve point of ,a system describes a trajectory, whose differential equations are, in canonical torm,
4i =·~1 ' Pi =  ~~ (i = l;~, ... "f); (~6)
the c3nonical coordinates and momenta are denoted, as usual, b~ qi' Pi and the Hamiltonian by E(q,p) (this is now the Ham1ltoni~ or the whole system)·, The anErgy. is a constant of the motion, and its value E definas a "surfacell in phase space
E(p,q) = E,'
(2.7)
in which the trajectory is entirely contained. ·WQ assume that all energy surfaces are confined to a finite domain II of phase space.
The value of snY macroscopic quantity pertaining to the system may be interpreted, from the atoc1stic point or view, as the average of a corresponding atomic qu~tity over the. interval ot tloe nec~s8ary for its mcasure~ent. If A(q,p) 1s the atomic quantity in question, such an average wduld have the form
to+T 1 r .
~.~ A [q(t),p(t)J dt,
o
..
.
~.. ,
•
where the Q1(t), Pi(t) represent the solution of the equntions of motion (26) for the conditions at the initlcl·instant to. The tine
• 1
25
intervai T o£ the measurement rooy be extremely short on a cacroscopic scsle and yet cover the passcgc at the systeJ.:1 through a great many
•
atomic configurations: it is then called IIphys1call~ intinitesimelll•
Now, we know by experience how a system behaves from the instant at. which all cannexions with the surroundings have been savered: atter
 a period of lldjUst/:ltnt, or urclaxat1onU, _ i t soon reaches a stationary
,. state in which it p?rslsts indefinitely. If we except the short per10d
~ of relaxation, any tina average will therefore have the same value over·
~.
any interval: the macroscopic value pertaining to the statl~nary state
of the systea cay thus be rcpresented by an averaga oxtending over a~
.' in:tini te time, an ide alized concept more amena ble to a ganeral analysis, :: tole, denote this averag_e over infinite time by
, to+T .
,(A)t"= lill _!..J A (q(t) ,pet) 1 dt;
.  T 9C1O ,T
, ·to '. 
(28)
, it is of course independent of the time to at which we suppose the
". sys tem. to ha va been isola ted from its environment. ' (
It is ,clear that the tioc average {A) t is _equivale~t to a .j'
.statistical average: for it cIJbod1es all the atomic configurations. actually taken by the system and i~ of course independont of the or
der in which they have been successivzly'takan. All we need to, know
~ is the relative frequencies ot occurrence of the various configurations of the systen in the course of its evolution: these will correspond to a definite statistical dlstritution on the energy surface, which will cocpletely characterize the equi1ibri~ state of the isolatecsystcn.' Siric~ this distribution is stationary, its density p(q,p) per alerlcnt of area do will not contain the time ,expliCitly and will thus satisfy the. condition
ft· t j><q,p) dcr (q,p) } q 'Ii q(t) = 0,
~. _ . ,; pet)
(29)
where the q(t), pet) 'again represent the trajectory •
•
In order to detercine the equilibrium distribution, some know
" ledge is nceded of the topological character at the trajectory. For
·26
. . :
. ... ..
 .... ~ .
. ,
~' the very large systens which interest us, one will expect that owing
.  , , "  .
c to the conplexity ot the interactions between the ato~ practically
" • I" •  •
'f" every contiguratloh' cc,mpatlblc with the external condi tlons will occur'
~ 800110r or Ie tcr • The " s1aple s t as. $U!Dpt iOD . w~ can [Jake? to, cxp~e S ~ ~h1S " .'. ~. intuitive surmise is that the trajectory passe's through ev::.ry,point ot '. ~, the energy surface:.', thiS, assuopt1on 'was adopticd by !!Oltzoann as tho "
r basis· ot : his argum,:nt' and he called  it  the erg cdi c hypothesis.. S tr1ctly'
~J speaking, one "should, :Dct introd1:lce any hypothesis in ~dition to the
~, eque t Lone ot motion;' ,~ut .1n':BoltzIi!ann~ s time the .fJathe.tlaticcl methods
f, necessary for stu~+ng the. topological prope~t1es of' ~r~j~ctor~es:h~d
C not yet bean deyalap~d and SODe .gucas had to~be .:riskad .. in order to, make,. ',> ~ny progre ss at' '~li. :::,Bol t zc6iin' _ ~nd~~taitMell ,~n gard~ d :the c; rgcdic . hYpo~ "
h" . . . .: o. .+    .• ~'.' •• " . ":~ '/ ... :" r .. ".1 , .. _. : i ; ~ . .• . ~
i.. thesis as 0, naturc:'l generalization of the Sto~sza~lansntz of' the kinetic
~" theory of go se s. ! ~~>~a scribing t~e: re~~ t :~f~· t~~:L ;"91S'o~~~:r~yi~ . ~~1o~s
~( .. at atomic systans~. . "," .. ",;
;;, ' If' we, a ceept ,the ergodic hype. the sis, it is easy to sec that
~:' the condition, (.::!9) uniquely d,:t~rmncs_ the distr,ibution dens~ty ,.o(q,p)
~,. ·on the energy surface. To this and~ we start frClil the iiJ.portant .prc 
t: party of' dynacical systcns, exp:rijs~ec:1 by Lio_uville IS thcoreo, viz. that ~ t'pe volum!? of' any donaiD,D_ .af ~ase 'space (dcfin~t;1 in or~n~rY' cu~11dean
~L ~~~s~c ': ai( ,'J.: , d:c:., '~.~ tk ~~.' ii d41: •• ~~qf':":, aPl ~;. ~~p;) :i~: irlv~i1knt" '~i th
, ~ .. ~,~ ~'.~ ,~ . ~ .: D·"~ '!'":  ~ ~"  . ~......  ... ~ .., , •• ..i."... " ..... 1 r. ., ~~ ... I '1 "'. It •
;. rasj)c ct to the" mbtioD:" ::.'  1 ,'~. '.... .; .,.,' " ...  '" '~'., .: _'" ,.l. '.~. .• l.Y
;,'," ', ._.; ....•.. " .. ,.,. :r~~d:··1J., ,'. ~
r.: '.:,. '.1 ~. ,",.j. ~ .. ""~ ... :=. .' : ........ :~ ........ 'dt.:: () ....
~,;: ,. . . :;, :9.;, '<D',· :.: '
.~
~:.'~ ~.
~~' .~ ..... ~; ••..•.•• . J.~ .......  •. ~l ': ..• :.~;.. I._~:". ~ .....
J' ~p v~~t':l~ .. o~\., ~hc~.Hatd_~~~!i9ia~: R,q~e~.~:?~~~G!~>.·: .~,~~~ ~9al~.!.~\;a'!p. ... tl!¥. ~,oof ;, Of .. tlUs ~ thccrc:lfor our la:ter .. systeaat1c, treat!2cnt;, let uS .. now .pursue .
I .~h~·'·~~~~~~~. ~APP~~g ;~'Lf~vi;ti~ ~ ~.: the~·7~p1:~·~·e="·5!·~:_et~·~~~;~~~ ~~~~~ .. ~f" •
~:' th~ ~,I ~hc:il~'. ': of. phase s'l?~cCf:'~ c.c:up~;se.4 ,\::~t\f~€n ;~c en~:t:gy ~~t~ccs:, ~ '.~ 1 ~,.E+~:".'w~ ~ ob~~i~i~h.e ~nya~;~!lt, GtCD~n~ ,9~, ~: .: ~, ~h£FC ,;~ ,1~.: the:. " ,:t _." thickna SS~! .. of , ~hc . sh<tl.1 .. counted along. the. no~nal to the, e l€Elent of .
~ .... ... . .. •.. _.... ....  _.I. • ..... ~~_. .' _ ,,.., _ ~.~ ........ _ r .... ., ....~, ~ r ....  • •
i . a~c a. _ ~cr ,_ ~hi~ , .. ~~~C~?~!':_",<::~~" ': be ': 1?~;Lc~l? ~~d in te ~~s~~f' ,:the )I~ci~~op:
t ? from .. ,< < "../t'(_ J(!~E ):~< '(aE~ZJ;' . _ ,' ':  )'  ,. _: .. t.: '.
~ "', ,r~' .';' ," ," ', = . .; r' & ,', , "'+ ..  'J' f£ 'v\ q, po .;' , ....  .' ~
J _. ,.; :".;;~;: !... do "i #Cli ,api
;~.
.,'
.'
2.7
the ttv'eloc1ty" IE ('til + Pi2.) at the point (q,p) of the trajectory. " stetionary distributions on .tho energy surface have therefore th( g~n
. era1 form F(E,Il, •• , .. , I2.fZ) x v{q, p) ld e, where F is a runct ron of
, a,set· of (2f1) timeindepandent integrals of the motion. Now, accord.' ing to the ergodic hypothesis, the trajectory 1s completely character
i 1zad by the Gbergy/1ntegr31, and the lunation F accordingly reduces ·to
: a function of E only; this 18 just'a tiol'calitat1on factor, equal to ~the inverse of the integral
;" t (E) = f 4T ·
E
.,
We thus ccncludc that the equilibrium state of an isolated system. is de·
fined by the ergodic distribution 1a!
;a(q,p)dcr :: 1 . d.,.... (30)
• ;
2: (E) v(q,p) .the corresponding ergodic average of any qu~ntity A(q,p) is
(A> E iii 1 r A(q,p) dG'"" •
~(E) ~ v{q,p)
(31)
The eq~ivalcncc between' the tine average (A)t averc.ge < A ) E is now readily verified. We wr1 to
(A)t =.·«A)t >E
and.~he ergodic
. i
•
~ : ' . ,
and, taking account ot (29), wa intercha,nge the or~~r of the averaging processes on the righthand sid~; this gives
<!~ ) 1: = «A > E) t = (A > E •
This equ31ity (A) t = (A) E' known as .. the aaan ergodiC thcorcn. is indeed the true basis of statistical thermodynamics. But the ergodic hypothesis, by which Bbltznann tried to es~abllsh it, wos later preyed to be wrong: a correct. formulation of the ergodic theoren and the·con. ditions for its volidity' .. rcquires the ideas of the acdcrn ttH~cry of
sets and cf integration due to Borel and Lebesgue. This is a quite recent dcvelop3ent, which we shall discuss in due course. For the aocent, there 1s only one important aspect of the ergodic theorem we want'
~8
to stress, viz. its relation to the problem of irreversibility.
In the preceding argument, we OL~sidered a system whose int
. "
·tial state was described bya welldefined phase; it would have been
more realistic to start"frow some distribution of initial phases in an energy shell, expressing the incompleteness of the det~rm1nation of the dynamical state o~the macroscopic system. The ergo~ic theorem states
~that any such" arbitrary phase distribution will in the course ot time .spread evenly over the whole energy shell, so as to take the limiting form of tBe ergodic distribution; in other words, the motion of the ,phases 1n the sh~ll can be. compared with a mixing 'process. We here
faca in its most general form th~ paradox of an essent1.ally reversible mechanism leading to an ~rreversible situation; , but we have a~so in hand the means of solving this paradox. If we analyse at any instant the distribution arising from any initial one by letting the size of the elements of phase space decrease indefinitely, we shall find that
. .
a given infin~tesirnal elemant is either occupied or empty, and the frac· _tten of ~ccupicd elements, by Liouville's theorem, will always remain 'the same: from this point of view, there is thus no mixing at all, but.
.
. 8 perfectly reverSible evolution of the distribution. If however, we
f'ix our attention upon, an element of phase' space of arbitrary but
finite size, the density ot distribution in this element wili tend asymptotically to a constant vulue: QS Ehrenfest expressed it, the irreversible behaviour pertains to a coarse distribution·in phase. The degree of' II ccareenass" i.e., the size of the cells of' phase space with respect to which the _distribution is defined, is arb1trarYi but it 15 essential that some finite subdivision into cells be assumed. It is this callstructure which r.:?prescnts the lack of'def'1n1t;l.on of the
! •
state ot the systCr.l necessary f'orthe application of a statistical mode
of' description of its behaviour. From the mathematical point of view,
we need not specity the' si,ze of the cel~s and we may even ulticat~ly regard them as infinitesimal for the purpose of' pr~ctical calculation; but the limiting proccssor letting t:.\: cellsize tend to' zero ;JUst ther be carried out"after the process of letting the tlwc of evolution of thE
, .'
systC.:l tend to infinity. If' vc should reverse the ·order ot the tvo .
l1~~1ng ,processes, we should obtain the reversible behaviour of a UtinE' distribution.
•
11. The Canonical Distribution.
The next essential qUGstion that confronts us 1s the fol1ow
'lng. The systGms we usually de[~l with are not isolated, but can ,freely exchange energy wi~h their surroundings. In particular, in order to assign to any system S a definite temperature, we must keep it in ther~ mal contact with J 'thernostat ~ i.e., SOr.lG system with such a large .
. ,
hoat capacity th~t the exchange of heat between it and tho system S
does not sensibly ~lter its ow temperature; in other words, the thermostat acts as a practically infinlt2 reservoir of heat at a definite temperature. There is thus a ,ce;tain oPPosition, or complementarity,be tween an isolated system and a system of given temperature: while the
_ former is Characterized by a definite value of the energy, in the lat~ ter the energy is essentially undetermined; the concepts of (dynamical) energy and temperature are in this sense mutually exclusive. Nevertheless, the statistical law of equilibrium distribution of a system
S of given tEmperature "is, of course, itlplicitly ccntained in thE ergodic theorem: for we may treat the total system S + «D. for~d by' the given system and the thermostat as an isolated system.
There remains, however, the practical problem of reducing the ergodic . distribution of the SYStciU S + ® to a more managcaplc form: for it involves an integration ovarthe dynamic31 variables of the thermostat which is irrelevant to the physical situation'; the only fWlction of the thei"tnostot is to fix the value of a single oacrcscopic paramet.'~r, viz. the temperature. Boltzcanl''l,showed how the elimination of the thcrmcstat variables could be effected in a simple and general way, leaving an expression for the distribution denSity of the system S in tcrml:l of the variables of t'hat system only. This density is new
, ,
defined'for each voluoc clc~ent d~ of the phasc space of the system
S, in accordonce with the indeternination of thc.energy of the systeo. We quote it now without proof, reserving its dcrivotion an4 closer d18. cussion for later:
\P(P;q)' d~ = 1. epE(q,p;a) d~,
Z· (32)
With Z = f efm(q,pja)dW.
30
In this formula, the thermostat is represented by the parameter ~, which is the inverse of the ,absolute temperature e in dynacicol units; the occurrence of other cxt~rnal parameters in the Hamiltonian has boa) indicated. explicitly. The formal resemblance with Boltzmann's earlier result for tho distribution of' the single. molecules or a gas is not surprising: for ~ha gas itself' plays the role of thermostat with respect to each of its molecules •• The law (32), however, 1s of' muCh wider vulidity: it applies to any dynamical system of a large nwnbcr
 of degrees of' frcedoc. We shall denote it by the name canonical £!!tribution, which was given to it by Gibbs.
The astablishraent of tlte canonical distribution ccoplctes the stotistical onalysis of ~ynamical systems, at any rate insofar as theSE systems are c1osed, i.e. have a fixed nuober or constituents.. The application of' thermodynamics to chemical processes, which becanc of great inportance with the growth of' che~cal industry in the last dec: odes of the XIXth century, required the reooval of this last restriction. The extension of both thercodynamical and statistical thec~y
to .~ systems is the great ,work of Gibbs, who introduced the concept  of cheillcal potential an~ carried Boltzaann's ideasto their'last de
,gree of generality. Yet the developmen~ or stati.stical thermodynamicsin this first phase, however brilliant an '8chievcncnt, was far frOD being a complete success: 0 dark .cloud hung over it, portending as it
+ we~e the radical transformation which atomic theory had still to undar~ go.
The root of the difficulties is the thaorC'!,ta of egt4Jt£g"_~,i t:!&!!. . .2! energy, which 1s a direct ccnsequence of the canonical distribution law, and holds .therefore in a much wider sense than the special example we net above when discussing atomic collisions. In fact, the kinetic energy, being a quadratic and homogeneous functi,on of theJ:1·,~,mentum coordinates, can be: written in the fora
•
,
• ~~.. _... .. ..... ....... ... _ ,.. ....... _ .... ..'1liiiiiii0 .....
•
*Hcncc also the possibility, nt first sight paradoxical, ot obtaining
an equilibrium distribution for a gas at a definite temperature by treating it as an isolated system.
31
and ter any onc term of this sum we have
I aK ~  1 I" t3 (eI3K, d
P1 aP1 e dP1   T J Pi clP1 P1 '
= +.+ J eJ3KdP1;
the canonical ave~ag~ thus beccues
<1 P aK 2" 1~
\ = ~ 8 :
Ie 2.
this means that each degree of' trecdcQ of' kinetic energy contribute~
. ~e to the average energy of the ·system. By a s1m11orcalculation, , the SB3C conclusion con .be extended tc those degrees of freed~ of potential energy with respect to which the potential function is 'quadratic and ~ocrogencous: an icportant example is the case'of harmonic oscillations around static equilibrium ccnfigurations, such as occur in crystals.'
In terms ~f dirGctly observable quantities, the law of equipartition yields definite predictions regarding specific heats. The translational motions cf the Ilol=cules in gases, e.g., give rise to
a contribution of ~ R per '!lole to ,the specific heat of the eos; the oscUlatcry motions just ncntacned in a crystal c'ontribute 3R pcr ~lc to the specific heat, It should be stressed that in predictions of this kind atomc theory g.oes further than macroscopic thcrmodyna.ncs: the latter establishes certain r~lntions between physical quantities, but the v3lues of S041(; ,of' them aus t be obtained directly froQ experience; atonic theory, on the other 'hand, suppl1esa dynamical interpre. _ totion for any physica'! quantity, from ,which the value ot this quantity can be conputed, at least in principle. The coupariscn of' the above stateaents ccnccrning specific heats with the empirical facts
is therefore of Breat interest ·to assess the value of the stntistlcnl
•
theory.
Now, such a coupard aon reveals a very c?aplcx state' 'cf affairs 'kt first siBht, the situation looks rather enccuraging. Hcnoatoc1c gases do have £l speci~1c heat (at ·cc.nst~nt volU!!le) of very nearly ~ R per nole in a wide range '01' tenperatures and pressures; as to the theo!
,z
ret1cal figure of 3R par ~ole for the vibrational specific heat or crystals, it corresponds quite nicely to th~ law of Dulong and Patit. However, the rapid decrease of the specific heat in both cases when
. the temperature becomes sufficiently low is 10 sharp contradiction with the indcpendc~ce at temperature indicated by the theory. Moreover the evidence from opt1c~l spectra shows· that the atees and aolecules
are complex structures with a ccnsiderable nunbar ot internal degrees . .. or freedom: the corresponding ccntr1but1ons to the specific heat which the theory leads ,us to expect a~c strikingly absent up to the highest 'taaparaturcs at which measurements have been made. We arc thus faced, undoubtedly, with quite a serious failure of the equipartition theorcc,
•
 but it 1s not a cOQplete failure. It rather looks as if the various
degrees of freedom cf the atc1:lic systems were graduully "frozen 1n'
as the tccperature decreases, so that thG law or cquipartiticn would only hav9_ssyoptot1c validity. The tcoperature at which the "freezing 1~1 appears would seem to depend on sane characteristic frequency of the mode of aotion conSidered. The internal atomc Q(tions of high frequency (cc~responding to" the frequencies of the acco~panying light radiations) freeze in.at ~uch higher temperatures. than the slower oscillatory aot1onsof atoms in molecules and crystals.
33
l;.; .. \t·~·~ t: ,:. ;:~~L~J ~i:; f~Url!S l~~~·~.t.: 13 I B L i 0 ... E C ,~
....... ,v.... H·~·.;_._
ctASSICAt ST~TISTICS
I.
I '
General Properties of Mechanical Systems
1_ Uotion as isomorphism in pha:sesoace. vie consider a mechan1 ~al system, the state of which 1s described by ! generalised coordinates  ql and their conjugate momenta Pie. The set (Qi,P1) i = 1, _ •••• ,f de~ fines a point P in a 2f'diLlcns1onalphase space, to which we assign tor ' ~ a reason soon to become apparent, a euclidean metric. The measure of an "~elecent of' phasespace is accordingly defined' as 1 ts eucl1deanvoluoc
r ::I dql·· .dqf dPJ.  • .dPt·
The succession of' states. occupied by the system in the course
or time for~ a curve.or trajectory in phase space, whose paramctr~c equa~_ t10ns can be written in Hamiltonian torm
qi = ; {i ' Pi =  i'~i (1)
with the help of the Haoiltanlan H(q;p). Tho system is assuced to bc conservative, i.e. H(q;p) = E 1s a constant of the notion.. In other ~,words, the trajectory detercined by the initial phase Po is contained ,1n the energy surface H{P) = E z H(P) passing through P. We ass~
. .00
that all the energy surfaces of cur system are contained within a t1n~
ita dorutin of' phase space: physically, this sitlply ocansthat no single coordinate or oomentum ever becomes 1n!inite. This assumption therefore, is a quite natural ona to make: it is f'undamental for the validity ot the whole ergodic theory.
It 1s convenient to vizualize the. motion, i.e. the passage troo
a phase Po to a phase Pt as a transformation ~f' phase space into itsolf. • The transrorQ3~ians Po+Pt obviously form a continuous group, of parameter t, whose infinitesimal transforoation is just given by the HaQiltoninn ~quat1ons (1). in other words, the notion is an autocorphiso
of' phase space. Integrals, like the Hamiltonian, are 1nvar1ants of'
. .
34
this group; the corresponding surfaces, like the energy surface, arc invariant domains.
I
2. Liouville's thcore~. For a statistical descript10n of a
stationary state ,of the syaten, we want Q measure in phasespace which 1s invariant for the Dotion group. The oeesure d~ introducod above has this property: Ithis is the statement ot Liouville's theorec. The
~aost striking way of proving this theorcc makes use' of an analogy bc
: tween the motion in phasespace regarded a~ a euclidean space and a hydro dynamical ot flow, whose velocity ~ is directly given at any point by the Hac11tcnian equations (1). The theorea tollows troQ the observation that this flow satisfies the' condition ot incoapress1bi11ty
..
. div v = 0, i.e.
r ~~. _." .
2:(~+~Pi)=o. 1=1 Qq1 ap1
This means in fact that any doonin Do of phase space will be transformed by the aotion into a,donain Dt at porhaps quite different shape, but of the SQtlC vo1uc.c.
From the invcr1ant tlcasurc df' in the 2tdimensional phase space it is easy to derive an invariant ncasure on t~e (2f1) diocnaional "energy surface. To.this end, consider an energyShell i.e., the.space betwccn two neighbouring energy surface E, E + dE. Let dl:, be the Euclidcon ~casure ot go eleoent ot the surtace R(P) = E, and dn an e1e
. Ilont of' length along the direction norcal to the surf'ace clement d"l: • The invariantmeasure d~E on the energy surtace is then:
. dr E = ~ = dE~ ~
' .
i.e. the euclidean element de is weighted with the roctor ·dn/dE. The
latta is just 1 grad H{P) 11 = { \ [;~/ + (~~/~J) 1/2
it is numerically equal to tne inverse of' the modulus ot the veloCity ~at point P. In the f'ollowing, we ~hall always usc the invariant . neasurc dr E' and we shull drop the index E when no contusion 1s to be tOBrad.
35
~ .~ ...
r, :;. Physically equivalent phases and uniforo phase functions •
...
tb the preceding sections we hove singled out the energy integral as
~ ,
: defining in 2t~diwensional space 0 nanitold of (zrl) dimensions which
~,completely contains a giVen trD.jGctor'y. It is necessary to explain
:;',thC! physico.i reason for dOU18 SO and not pursuing the reduction or. the tnuober of didchsion~ ,of thi§. mdnitolijany fUrther·with the heip or the t:_othcr timeindependent in~egrQls ~ Fol" this _ purpol3~; a cioser cons1cie1a.. •
~ .
~~t1ori ot the phys1ccl interpretAtion of the forna11sm is needed.

~:' It generally occurs that the same physical state of the systCQ
•
~is represented by acre than one phase. An cxaople is offered by an
t gular variables: phases in which ·the values _ of sene angUlar variables ~, differ ~y 27T describe the s.ame physical state. Another case, of less Fformal character and of fundamental importance in atomistic physics, is
that of systems consistlllg of identical elements. A phase 1s then defined by the sets of canonical coordinates pertaining to all the elements, enumerated in a certain order; two phases differing only by this order of enumeration are indistinguishable from the physical point of
view.
To express th1s situation, one calls the phases as just defined
specific phases, and one denotes by generic phase the set or all specific phases corresponding to the same values of all coordinates, taken in any order. A physical state of the system is thus described by a generic phase.
Phase functions representing physical quantities must have the "same value for all phases corresponding to the Same physical state ot "the system. Thus must, be periodic with,respect to any angular variable *,"and if they refer to a system of identical elements they must be symetrical with respect to permutations of the sets at variables pertaining to different elements. Phase functions possessing" this
property will be cal~ed uniform.
Now, it we consider u set of (2£1) independent lnt~grals not
,
r
~~~~~~~~~~
•
* Strictly speak~ng angular variables do not satisfy the finiteness conditlonenunciated above" (paragraph '1)". But they nevertheless can Mt ; used just on account of the periodicity property ot all phase functions of physical significano,. See the example discussed in section 4 below.
36
containing the time, which fix the shape of 0 trajectory, they will not in general be all uniform. We can assign the value of any unitorm integral, which means that we can control it by physical means; but it has no physical meaning to assign a value to a nonuniform integral. The latter can therefore in any case not be used to reduce
. the manifold containing the trajectory. As regards the uniform integrals, the aSSign~nt of a definite value to anyone of them is purely • a matter of fixing the conditions· under which we wish to consider the system under investigation.
ThUS, we define an isolated system by fixing the value of the energy integral. We might wish to fix the value of, the momentum and
•
angular momentum integrals: this would mean that we consider the sys
tem as free to move in empty space. In this case, the number of dimensions of the surface on which 'the trajectory is C(\.t")l!!:t:!.rcd would be reduced by 6 mora Uni~s~ It will be more in accordance with U~l1!\l. conditions, however, to leave these integrals indetermi~ed by imagining the system in contact with an infinitely heavy body, with wh~ch the elements of the system interact elastically. We shall. therefore
.. '
base the following argument on tho .. considaration o:t" the ener~y surface;
nothing essential would be changed, ot course, by "';he assumption ot a "surfacell of' a somm'That smaller number of dimcnsiOl~s, corresponding
to the aSSignment of fixed values to other uniform integrals as well.
4. Example. The above considerations cay be. illustrated by
the simple e~mple of a system of two uncoupled rotations aro~'d fixed axes. Let:,Us take as coord1nate~ the a2imuths ql' Cli· ani assune for simp11ci ty the t,'IO momen ts of' inertia e qual to uni ty ; the cor. ~~ '.'.go te momenta PJ.' P2' which :represent the angular m.omentaar.our.~' the exes are unif'orm integrals, to which we as Sign the values W:L ,w 2. (..;urn.r:: l"i~ cally equal, in our case, to the angular VGlocities of rotatiol";) i· this also fixes the energy E = ~ (Wi + w?2 ~. .The !ilBllifold centlini.'1g.
the trajectory .thus red':1:cos to the plane Yol.' qz and the t:r:'fljactory,
1s iif straight line in this plane. .• ~.
L
t:·
. 37
The plane is divided by the lines
ql = m.2vt q2 = n.2u (m,n integers) into a chcquer of squares all conta1n~' lng phases physically equivalent to
fl' '
r
"" +o,,...,...,...r_4000000i·.....
t",.' each otherit For all ecmput at Lon or
~~~~~~~~r~l aVerages or physically significant
L phasef':u,not1ons, we Clay accordingly
, .
;t" strict the manifold containing the
r trajectory to a single square, by
~. transferring to this square all the
~. ~egmcnts of the trajectory contained, in the ether squares. The equivalent
..
r'": trajectory thus obtained cells,lsts of a set of parallel segments with
, til the single square; this se t 1s f1ni te or infinl te ace ording as the
ql:lantltY. ~3/W2 is rational orirratlonal.
re
",
.., .
. ~ ..
, .
Now, we may choose 'as a third time1ndependent integral the
function
. qabi)l ~ qi w 2 =' qaPl  qlPZ :s M
integral is not uniform. In fact, it takes a different value on each ot the finite or infinite set of segments composing the trajectory:
~, tor on the segment originally in the square containing the phase
(qi + m • Zu, q2° ~ n • Z~), the value of M differs from'MQ by
211 (nOll.  mwZ).
5. Uetrical indecomposobility of energy surface. A question
.,ot primary iaportance for the estnhlishrlent of a statistical distribution on the energy surface is to characterize froa the metric~l point of ~iew the set of phases occupied by the system in the course of time, i.e. to know the measure of this set on the energy surface. Our simple example (4) sugeests that; apart from exceptional cases arising from some Udegeneracy" (W1/t..)Z rational), any trajectory will in a certain sense tlf111" the whole energy surface. We must try to give to this intuitive description ot the physical situation a
38
rigorous mathematical formulation.
Bolt~mann and Maxwell assumed that the trajectory actu~lly passes through every point of the energy surface: this 1s the famous ergodic hypothesis. It is easy, however, by the modern methods of thG theory of sets, to disprove this hypothesis (Rosenthal and Plancherel). Consider a neighbourhood S of a phase occupied at some time by the system; it can be chosen sufficiently small to prevent the trajectory from remaining inside it at all times. Then it is clear that the part
.. of the trajectory 1nside the neighbourhoOd will consist of a set of separate segments. Moreover, the time intcrval·s during which the system. 15 inside S will form a succession of separate finite segments of the time axis. Accor~ing to a wellknown pr·operty of sets of points: the set of such time intervals 1s enumerable. The segments of the trajectory inside S thus form an enumerable set, whose measure on the energy surface is zero.
A correct formulation of the i1ergodic" situation is obtained . froti the consideration of the autOl!lorphism defined by the motion of the system. The set of phases through which a· trajectory passes is
. clearly an invariant set with respect to this group, and we are concerned with its measure on the energy surface. Let us call the energy
t .
surfaca ~icallY indecomposable (and the ~roup metrically transitive)
it it cannot be expressed as the sum of two invariant sets both of ·positive measure. This oeans, then, that the set of phases forming
a trajectory either 1s of aeasure zero (this is an exceptional case, such as that of periodic motioD; with a closed ~rajec~ory), or has
the sane measure as the whole energy surface.
At first sight, the condition· of metrical indccooposabllity Of. the energy surface would seem impossible to. f'ulf'ill. Let us in tact consider any timeindependent integral I(P), different fron the Hamiltonian;. t~c phasefunction I(P) therefore cannot have the saoe value over the whole energy surface. But then it 1s always possible
. .
to find a nuaber I such that the invariant sets of phases· for which
I(P» I and I(P).~I, ~espectively, are both of positive measure*. At.
•
..
* For the proof, see Khinch1n, p.30, footnote.·
39
this point it is necessary to remem1:erthat ther~ may be different phases corresponding to the same physidal state of the system; we might then have an ihvariant subdivision of the ~ne~gy surface into parts ot positive measute, sueh that no two phasas of ahy part are physically ~quivalent (e:.g~ the chequez of square s in our example, paragraph 4) sueh a SUbdiv1Sion'/whi1d formally violating the condition of metrical indecomposability as formulated above, would nevertheless represent, physically, an Uergodlcil s1tuation.
We must accordingly modify the definition of metrically indecomposable sets So as to cope with this case. \'1e call an invariant subdivision of the energy surface Jnto.two parts of positive measure essential when all physically equivalent phases belong to the same part. The surface will then be metrically indeoomposable in the physical sense when it does not allow of any es.sential Subdivision. This modification will not affect the above argument in respect to Wl1torm integrals: these will bring about an essential subdivision ot the energy sutrace. But we cannot say anything about the effect of nonuniform integrals; in fact, in the example of paragraph 4, it can be shown that the nonWliform integral M in the general case (Voll/WZ irrational) does not disturb the metrical'indecomposability at the (reduced)energy surface.
The next question would be, how from the structure of the Hamiltonian could one draw conclusions regarding tae metrical indecomposab11it of the corresponding reduced manifold. This problem is not solved, how
III
ever; inthis sense, the assumption of the metrical indecomposability ot
the reduced manifold remains an hypothesis. The progress with respect to the original ergodic hypothesis lies in its precise .mathematical for~ mulation.
As we have just seen, we can only assume metrical indecomposab11ity for the manifold reduced by taking account ot all uniform inte

grals. In practice, however, the uniform integrals distinct from the
energy will usually have the same value over a very large part of the energy, surface, and it will be possible to neg_lect the dOillainsof ver~" ·small measure in which they differ trom this dominant value. We may
•
. 
therefore still restrict ourselves to the consideration of the energy
surface and speak, in this approximate sense, of its metrical indecom
posability. ~
'.
•
40 ..
, 
•
; .
6. Poincare's theorem. From the qiscllssion of the "shapell of
the trajectories we now pass to the consideration of the way in which
,
they are followed by the system in the course of time. This question
is dominated by a theorem enunciated by Poincare, and proved rigorousl~ much later by Caratheodory: for almost all trajectories, the system
,returns after a r~ficiently long time arbitrar~ly near to its initia: phase. An es sentlal condi t ion for the valid1 ty ot Poincare I s the are ~ . is the finiteness ofthe energy surface conto1ning the trajectories.
The proof consists in showing that the set of exceptionul trajectories, i.~. the set of those phases near which the systemnever returns, is of measure 'zero. We must first tormulate in a precise wal what we mean by the "~cturn" o:f the system near some phase Po.Let us choose some time intcn"val ~ and consider the sequence (p) of phases
. ,
Po,P1, Pz · .,.' successively occupied by the system at tines toto+t; to+2 C, ....• We shall then ~ay that the system returns near Po it 'every neighbourhood Op ot Po, however so.all, corrtadns at least a point ot the sequence (~). The ~xc€ptional phases will thus be those,
. '. ~
for which a neighbourhood can be "found containing no point of the se:
quence (p). vie must pr.ove that the set of 'exceptional phases is of aeasuze zero.
To this end, we cover the energy surface with a net,' dividing
. it into· intervals Ui, and so finemeshed that for every point P and every neighbourhood 6 p o:f P, thc,re is at least one ·interval Uk containing p and contained within ~p: this means that we must actually have an infinitely fiae mesh, and an enumerable infinity of intervals Ui• Let now Di be the set of phases of U1 whose timesequence (p)
, never returns to Ui or its boundary; th~ set Di is therefore open and ' accordingly measurable. Moreover, the_ 'sum D = ~ Di' is the set o:f
exceptiona~ phases: in fact, every point of D is clearly an excaption a1 phase, and conversely, it will readily be seen that any exceptional
•
phase must belong to one of the sets Di'.  .
The mefisure ~f D being the sum, of the measures 'of the Di' s , the
, theorem will be established ·if we 'show th~t each e~ceptional set Di is _, 9£ ~easure zero. Let us consider the sequ~nce of trans:forms Di(n) of :
41
the' se t Di at the times to+ n'G: no two of these sets Di(n) can overlap, since otherwise they llould contain nonexceptional phases. The
: measure of th~ total set E Di W is therefore the ,sum of the measures ;, n(D1(n», which by Liouvil~els theorem are, all equal tol. (Di). But
': O\oIing to our as S umpti on that the tre je ctory is con ta iIl,? d wi thin a £1 'nite domain ot phase space, the measure ,Q (ir Di(;:J.) ) must be bounded:
: this can only occur if n (Di) = 0.  .
. ,
7. Th~ return t~!Je. A consequence of Poincare's theorem is
, that almost any state of a dynamical system will recur to any given approximation in tae course of time, and this recurrence will repeat itself an infinity of 'times. The evolution of the system will thus e~i_ bit a II quas iperiodicll behaviour. The II quasi  periocilt, or ave rage time ~" ot recurrence ot a given phase to. a specified approximation will, however, not only depend on the demanded accuracy ot reproduction of the
original phase, but above all on the number of degrees of freedom of the system, The sensitiveness olthe latter dependence accounts for the tact that the recurrence of a phase of a system of macroscopic
~
diraens10ns is never observed, "Ihereas the quas1periodicity expressed
, ,
by Poincare's theorem is the rule for smaller systems. A simple example
will illustrate this situation and give an idea of the orders; of. magnitude involved.
Consider a vessel containing a~ molecules of an ideal gas under ~'normal conditions of temperature and pressur~. The probability of finding an excess of molecules, in one )lal,f of the vessel comprised between
5 • loT and (~+ dS) N is, asymptotically 2.
. (AS)
11" (S ) d S :::!. 2_ e  d ~ , ~r "
.' {"TiN 2
ss: J ~ e5  N d 6 ,
, .
, if lIe assume that this inequality has arisen in the course of random :: fluctuations • Now,. if it as the average time interval needed for the " relative excess $ to change by I d tS I, the probability of occurrence
of an excess in the interval (0, ~ + d6) in the course of the evolu
; t1an ot the system may be expressed'as the traction dt/T of the average
•
.. , LJ,I::
time T be tween two successive occurrences of th(:! excess £. The ergodic theorem, as we shall see in the next cnapter , a11o\,ls us to equate the tllo probabilities just considered and in this way obtain an esti
~ mate of the recur~ence time T for the excess 6 :
I .
: The average time variation of the excess 6 _ 1s of the form
~. b "V e tit; ,
. ,.,here the relaxation time Z; is determined by the mechanics of the irreversible diffusion process leading to the est~bliShm.ent· of a uniform
· density. One has, theretore,
i ~J". _1
~ I ~ I  & ~
and . . ,:rr' 4 2 N 'l:
T =~rN e
If D is the diffusi~ coefficient, the order ot magnitUde oft; 1s related to the linear dlmens.ions L of the volume In''lhich the process takes. place by a formula of the type
L2 . 'G V> .. "" D
It the volume is of macroscopic dimensions, L~lcm, say, one has
_ r; ~ 1 sec and N ::::: 1019• Even an extremely small relative excess ot den .. sity,b ~lO6, would not recur s~ontaneouslY before times of the fan~ ·ta~tic order of magnitude T~elOsec. But ir the volumeis of micro
· scopd c dimensions, ee L ~105 cm, an excess of density of 1% ",ill.
· have a recurrence time of tb4 order of 109 sec .• '
43~
IJ. The Ergodic Theorems.
1. Existence of thEt time average. Let us ccnsf.de r the evolu
. tion of a system. starting at time to from a certain phase Po. As ex~ plained in the introduction, the measurement'ot·some macroscopic ~uan" tity corresponding Itto tile mt.croscopac phase function f(P) yields in
the first instance a time average t +T
~ F(i'o·;to,T) = if 0 r (Pt) dt, (1)
to
taken along the trajectory definea by the initial phase 'Po; in this ; formula , the symbol pt:denotes the s~t of canonical· coordinates of a
~ point of this trajectory,· expressed as functions of the time and the initial conditions Po, to. One is led to expect that ·this time of average has a limit for T ... 00, which is inde pendent of the initial conditions and represents the value of the macroscopic quantity for the system in its given stationary state. We must now discuss the mathematical justification of this surmise.
The first. step is described by Birkhoff's ergodictheorem, which states that the limit
F(Po) = 1i.m F(Po; to,T) T .... oo.
(~)
exists for almost all trajectories, and is ,independent of the initial time to. We first prove the theorem for the case that the time interval T varies by finiteincrem.ent~ of' duration·T, i.e. we take T= n 'Z:' and investigate the limits of the sequence
as ne ee , Let Pi? be an exceptional phase, i.e. such that Fn(Po; to> has no limit; this means that, the l~wer bound l (Po) and the upper bound F(Pc) are ·diffe~ent. We caD then choose a pair of members~,~ (~ < ~) between E(Po) and.P(Po)' i.e. such that
•
(4)
•
Nov, if the set D of exceptional phases \'1ere of positive, measure, it
44
is easily sccn* that ona could find a pair (~ ,~) for which the conditions (4) ere s~tisried for all the phases Po of a subset D* of positive measure. By showing that this last property ~ (D*» 0 is contradictory to the incqua~ity.O(.( p, we deduce that the set D of cxce p
~. tional phases 1s of 'measure zero.
The contradiction is elicited as follows. Let us consider the
sequence of times tk,= to+ k fj and the corresponding phases Pk = Pt
~ ( k being an integer of any sign), we def~ne k
 rtk+1
~ fk(P~) = ~Jtk f(Pt) dt (5)
: the time avcrage over the interval (tk, tk+f). By a change of origin
of the times, we sec that •
fk(Po) = fo' (Pk) , (6)
is expressed as nl
= n! ~ fk (Po> " k=O
and if we integrate this over any set of phases Do(n) , we get
, ! (n) Fn (Po; to) df = r: f (n) 1'k (P~) 'r = E J (n)1'o(P)~ ,.
~ ~O ~ ~O ~
where ~(n) is the transform 'Of the set Do(n) when Po~Pk. If now Ib(D) is a subset of D* such that for any P~ OiDo(n) one has
 F (Po; to) >~, we obtain the inequal:ity
, n 1
b( 1'0 (P) d;O n) n () ·
Suppose further that the sets l1r(n) are nonoverlopping, and call
their sum n1
D(n) g' ~ ~(n) k=O
Since, by Liouville' s theorem, fA (~(n» c fA~ (Do(n», we ·then have
* For details, see Khinchin, p. 1927
45
r fo (F) d .f> ~ fl (D(n»,
"'D(n)
Now it can be sholm that such sums of nonoverlapping sets D(n) can actually be specified for each value of n in such a way that they together exhaust D*. We can then sum the inequalities of the type just derived for a~l n, and we get in the limit n~
I ro(p) df) ~ft (D*) •
~* .
A S~milar argument leads to ~*f'o(P) r <o<y. (D*)
,
•
and these two inequalities tnth"jU(D*» 0, do indeed contradict the
1nequali tyg( < p • " •
To complete the proof of Birkhoofffs theorem, we have to compare the average fo~ an arbitrary time interval T with that for the interval of the sequence n C nearest to T. We have
to+T to+T
I ];. r . f(Pt) dt .: ~ r f(Pt) dt 1'0,
T Jt 0 "Jto '
t +T 'to+nl',; to+·T
Je~oO ;(Pt) dt  iftiof'(Pt)dt'  ~{+!;t)dtl ~~I
and
with the notation (5). It is readily verified that
lim 1

n~cc n
almost everywhere. The proof is of the familiar type: one shows by
, "
making use of (6) and of Liouville's theorem, that the set of phases
Po for, which Ifn (Po) I} n £ 1s of measure zero. Hence, the time average F (Po; to,T) has a limit for T+oo almost ever~ihere.
Finally, it must "be shwon "that this limit is independent of
~ ~. '. ~J ;.\ .~ ._'
the initial ti~e to. We. have ,
, '. 1. tl ... .,. 1 ltl+T
11m ! ~ 11m =
T 0 T~to to
" t
46
_since the dif~erence of the last two expressions tends to zero like , (t1tO) IT. Further, the difference
c 1 Jt1+T 1 rt1+T _ 1 jtl
r T ,to  T tl T t.o
~also tends to zero, (which complete the proof.
...
2. Equivalence of time average and statistical average,
""
~he tiQe average whose existence is established by Birkhoft's theorem
.does not yet correspond, in general, to the physical notion of a macro.scopt c quantity attached to an isolated system, since it may still de,pend on the particular traj,ectory followed by the system and have difJ'erent values for different trajecto:";"ies even on the same energy surface. 'It is easy, hovever, to indicate a genera~ condition sufficient to ensure the constancy of all time ave~ages almost everywhere on the energy
, '
surface: ,it is the metricel indecomposability of this surface, at least
'in the physical sense , Fori£ the time average F(P) is not almost' every'where' cons t.arrt ," it is possible ,to find a value F of F(P) such that the ',~ondi tions F{P) > F and F(P), F define two invariant sets of positive ,measure, effecting an essential decomposition of the energy surface.
If the energy surface· is metrically indecomposable, the constant value of the time Bverage can be expressed as ~ statistical aver, ,.age over this surfac~. Roughly speaking, the trajectory along which 'the' time average is taken IIfil1sn the whole energy surface: the time ,average can thus be conSidered" if' ,re disregard the temporal succession .of the pha se s as an ave rage ove r the energy s uria ce, wi th a def'ini te .weighting of each surface element. This statistical ~leight turns out to have an extremely simple specification: if we use the invariant measure on the surface, the statistical distribution equivalent to the time average is uniform; in ~thcr vords, the amount ot time spent by the system in any region of the surface is proportional to tho inv~riant measureo! this region. This uniform .statlstic&i· d1stri~ution is called ergodic fu,d the fun~amental corollary ot Birkhoft's theorem which express this situation can be formulated as tollows:
If the energy surtace{1is metrically indecomposable, the
•
~.'
47
tima' average F of the phase function rep) is 'given by the ergodic aver
llge
(1)
In the first place, it is clear that the ergodic ave~ag~ of rep) 18 the  same a's that of the/time average,F(Po; to, T),over any finite time in
terval T; indeed
to+T
dfE to ~ (cit jJ~n);}Slt(Pt) dfE
in virtue of the
lnvar1ance  of1 the energysurta ce n ;
to+ . '
~ r 'lIE dt '; liE
'. T Jto "
Since, owing to the 'constancy of F,
and this 1s
...
F = 1., ~ F dfE'  fl (m J~
it remains to prove that
i
This 18 read11y*. ~een to be a consequence of the ergodic theorem F(Po; to,T) ..... F tor T~(D tor almost all Po.
The last step established in the phys1c~1 equivalence between time average and 'ergodic average 1s an obvious extension ot the preceding theorem;
If the energy surtace 11 is metrically indecomposable in ~ physical sense, .the time~ average F or any l¥llform phase 'function rep) is equal to its ergodi"c.average;
. F:: tiE (8)
 ,
* See Kinchin; p. 3132~
48
The equation (8) is valid for almost all phases of the energy surface. For the physical interpretation,.lt is therefore necessary
to introduce a further averaging in order to eliminate the possibility that the initial phase would happen to be an exceptional one. Instead of starting, from a ~elldetined initial phase., we thus aasume an initial distribution of IIcdpiesn of our system over a'finite domain o~ the
· energy surface. The time average F must then be averaged again over
_ this domain: any exceptional phases contained in it will not give any
· contribution to the average, and the result will again be expressed by
an equation perfectly similar to (8). The replacement of an initial .: phase by an in1 tisl 11 cella of finite (and ar.bi trary) size· has a great :'importance for the physlc~l interpretation of the theory. Before we : discuss it, however, we shall proceed to an e~tension of the ergodic
theory which first displays its full scope.
3. Hoptls ergOdiC theorem.
The restriction or the phases to an ene'rgysurrace is a· strong ideali_sation of an isolated system. It is marerealistic to allow a certain.
. 
· margin to the definition of the ene~y, i.e. to regard as possible all
trajectories contained within an :energy _shell, consisting of all energy surfaces in the energy interval (E,E + dE). The Grgodic avcr~ge over an ~nergy sheil is evensoaewhat simpler of expression than that on
'" t~e energy surface, since 1 t corresponds to a distribution 6f Wliform · density with respect to the Simpler measure in phas~space given by
 Liouville's theorem: denoting the energy shell by (E) and the corre
sponding ergodic average by r '. (E)' ~e have
fiE = 1 J rep) df' ~ fiE»). (E)
here ~ l(E)] = J df. represents the measure of the energy shell.
 (E) '.
The condition !O; the equivalence of time ave~age and ergodic average over an energ~ shell is not simply that all the energy surfaces within the shell be metrically indecomposable. An additional condition " 1s needed, involving mutual relationships between these surfaces. The
•
49
situation will be made clear by a simpl,e example. Consider a motion of a Single degree of freedom, defined by an angle variable
_ q = wt (mod 211") and the conjugate action variable f , which is a con
:' stant of the motion. We may assume that the value of f.flxes the energy:
in the "phase space!! (p,q,) the nenergy surface" is then the 11ne
", p = const., 00r ratl1er the segment 0", Q ~ 2 on this line. An energy ~shell will be a strip limited.by two such lines. While each energy : 11ne 1s metrically indecomposable, the situation with respect to an
W •
;_~ energy strip will be radically dift,erent according as, the lntegrai p
~ is or 1s not independent of ~. In the .first case, any in1 tia1 distribution will simply be displaced along the strip without any t.endency
p '  . p~
•
2:11' q
11 q
1r q
,
to uniform spreading. In the latter ~ase,the initial distribution,
, .
,  while remaining of the same total area, will sprea d out in the form
__ of a more and more intricate subdivision of' partial domains stretching over the whole energy strip: this illustrates the "mixingU mechanism by which the ergodic districtuion 1s esta'b11shed asymptotically.
The general analysis of this "mixingll over an energyshell  has been carried out by Hopf. It introduces two new elements into
the problem. In the first plac~, it is neccessary to give a precise .>,
mathematical definition of the mixing process; moreover, \ie must for
, mulate the addition~l condition which guarantees this process to take place. The limiting' process corresponding to the mixing over an energy shell Is somewhat weaker than the convergence almost everYl/here which • obtains for the time ave~age on an energy surface: it has the character'
" of U convergence in the meanu• To define this, let us start, at time
to = 0 r=?m ~ome arb~ttary distribution characterized by a denSity g(P);:"" the statistical av~~~ge of a phasefunction f(~) for this,distribution
, i "
1s given by .
50
.... ~ . it ..
(f ,g)  f rep) • g(P) dt,
I the integ:.:at1on i:eing extended to the whole phasespace (the limitation
. :~
~to an,energy shell is contained 1n the form of the density function
~g(P) • At tiille i, the distribution densi,ty has become g(1t> , i.e. t::e ;,". transform of g (P) 'tty the automorphisCl P~P t; this 1s sene function ot ;P and 1 which we denote by tt g(P). The ave~age ot t(P) at time i is r accordingly (f, C' tg). The 11m! ting form g(P) at the distribution den
....
rstty is then defined, in the sense of conve:.gence in the mean, by the
: cond1tion
~
T
11i:il 1 f
f 'too T 0
(10)
•
~ for any pair: of funct iens f (P), g (P) • This aeans , physica lly, that the time ave~age of th; statistical fluctuations of the phase average of r(p} around its limiting value (t,i) tends to zero as the time T 1n
_ creases indet1ni tely. "Such a formulation 1s all that is needed for
:. 'physica1 applico tion's •
, We must °noll state the condition for the existance Or.8 11ait
1' •• 1ng . dis tri but len i (P) satisfying (10); as our example shews,. this con
"
~. d1t1cn tlUSt express a relation between different energy sUI'.f'aces: in
··1 ts simplest .f'orm it must rete:: to a pa 1r of energy surfaces • It is
.. convenient, .for the' considej."'ation' of such.a pair, to i1dupl1catell in a
....
certain sense the phase space, i.e. to introduce a phase space whose
: coordinates consist of a paii' of sets ot coordinates of the original , phase space: this nell phase' space. may be described as the "direct ,pj."'oductil of the original one into itself. A pair of energy surfaces : of the orieina1 phase space thus appears as a single energy surface
of the product spa<?,.e. We are nov ina position to enuncaa te * Hopf 's theorel'J: if' every energy suZ'.f'ace of the original phase' space is metrl: cally 1ndecor:J.posable, and a1nost eve",'y energy surface ot: the product
:' space has the same proparty, eva:y distribution tends,:to a l1miting
~ distribution of unifoi'El densd ty in the sense of equation (10). This
, <'
, ~
f " 'I'
I
,  I.
·1 .•
J
,r,
,
!
• I.
i
.. ;,
• We shall not give the proof of Roptls theorem, since it doea not ~nvolve any essentially new r~f.it~e.
,1
theorem completes the foundation of the concept af ergodic d1str1bu . t1on, on which the whole structurG of statistical thercodynamics rests.
4. The ergodic d1~tricut1on. '
Lat us now discuss froo a more physical point of view the main
" features of ~he e:r~c.d1c districution and of the mixing process by \.thich ; · 1 t comes about. In c._nsidering this process l{e iDlCiledlately t ace the
~. paradox of an essentially revers1tle aechan1sm lead1ng to an 1rrevers1bl t situation; but we have also in hand the .means of solving this paradox,
'.  t
If we analyse at any instant the distribution arising from any initial
: one by lett1l1g the size of the elements. ot phase, space decrease in
: definitely, we shall find that a given infinitesimal element is either occupied or empty, and the' traction ot occupied elchlents, hy Liouville's theo:rem, \'1111 always remain the same: from this point of vie"" there
is thus no wixing at nll, but a perfectly reversible evolution of the ~ distritution. If hcwever, we fix our attention upon 'an element of
phase space at arbitrary but finite size, the denslty of distribution
, .
· in this element will tend asymptotically to a constant value: as
 Ehrenfest eApressed it, the irreversible behaviour pertains to a
· coarse distribution in phase. The degree of "coarsenessll, i.G. the . size of the cells ot phase space with respect to Yhich the districu
... i
tiOD 1s defined, is arbitrary; but it 1s essential that some finite
subdivision into cells be assumed. It is this cells,tructure which
,,_ r:presents the lack of definition of the state of the system necessary for the application of' a statisttcalmode of description 'of" its· cchavio~. From the mathematical point of View, we nee~ not "specify
the size of the cells and we may even ultimately regard them as infinitesimal for the purpose of' prac.tical calculation; but the limiting p~ocess of letting the cellsize tend to zero must then be carried out aftc:.· the process of letting the tit.lc of evolution of the system tend to infinity. If we should reve:;.·se the order of' the two limiting proc~sses, we should obtain the reversible behaviour of. a 1If':lnC:U distribution.
Tho next probl~m.concc~ns th~ time of relaxation of the mixing process, i.e. t,he avezage time which must elapse before any given
·
· ,.
,
,
'II
,
· .
"
distribution has practically reached its asyr:1pto.tic uniforfl1ty. At first sight, it liQuId seem that l"fC are landing into another paradox; our exa~'lJ.ple ot section 3 suggests that lie .i:lust expect the ~Jixing to
" proceed dU~:'ing a time sut'f'icient to allow the systCIJ to ze turn severa]. tiwes near phases previously occupied, ~n accordance lnth PoincarS's theorem. But the pccurrcnce ot such ~cpctit1ons ot initial configurations is just the kind vf bahevicur characteristic ot the pur~ly
" me chan! cal evolu t1eD of" the sy ste rn a s opposed to 1 ts s ta tis t1 cc 1, ir:reversible! evolution. The fantastic order of Cil~gnitulie c·f the times ot recurrence accounts for the f'nilure to observe such repetitions 1n usual cases; and if the mixing process should really have to involve them, the whole statisticsl 1nterpret~tion of irreversibility would collapse.
To clear up thesituation, it·must be remembered that the
usual sys;teus to which th€:"mccynomics is applied arc coapcscd of a
large number of identicdl eleuents: th~y arc eithar bedias built up
of atoms or redia ti?p fie ldsc on sis ting of proper os cilIa tions • The phys1cel states of such systams are represented by generic phases: the latter do not corre spond to sil1g1e II cells" of phase space but· to regions called ,lconstellat1onslf l:y Ehrenfest containing many cells and 1ntI'ic~ 1a:tbricated througb each other oV€i'each energy shell. The ceasure of a constellation, according to the ergodic theoI'\;i"Qs, gives
the r~12tive probac1l1ty ot occurrence cf the stnte represented by the corresponding g~ner1c phase in the course of time. states ve::.y near
ttl equilibrium will thus OCCUPY' a much 'larger pa1:'t of the energy 'shell than states deviating apprec1ably from the equilibriuc conditions.
Thus, if a trajectory starts from such an lIimprobablcu c·.·nstellation,
it ~dll soon enter into cere and more probable ones and when it will have reDched the equilibritio constellation, it will stay in it most
ot the t1me, ~lthouzh it \dll continually cross less probable ones,
appeai"in8 as "fluctuations" from the equ111briU!!l state.
...
The r2turn or·
I.
a 3ie:l'Y iuprobacle stats, i.e. the~urn of the trajectory to ti" v~r:f icprocable constall~t1cn, would require a tiuw of a fantastic order
•
of magnitude; but the time of relax~tion is defined as that needed
t9 reach the most probacls constellation, starting from a less pro
: p
~ ....
"'53
bable one; and this ~s extreuely short as so~n.?5 the number of ele· ments 1s rcascnab1y large: The importance of the cc,ns1dc.:at1on of generic phases for the consistency of tho statistical theory 1n all cases in ,,,hieh its application is slgnif1cont was, ct course, tully recognized coth by Gibbs and by Bcltzmann. But the mode 9t C!xposit1on adopted by Gibbs tohded to obscure this issue. Gibbs vas "'Torried by the d1screpanc1es cctwcan theory and €xpcr1mcnt, eSpGc1ally with re
~ spcct to tho consequanccs or the equ1part1t1on of energy for the spe ...
.
o.
7 eific haats; these d1screpancies seemed to be connected with the as
~.' signment cor the nuober of dcarees. at treedo.cl of the atomc systems;
. in the hope of throwing light cn the origin ot the difficulty, he accordingly set hi2self the task to derive as rigorously as po~sible all
 those statistical properties which apply to the most general mechanical sys te ms, without any res triction as' to the number of de g re es of fre cdca, ,
. Hence the emphasis in his book upon the derivat10n of such results, while the discussion of systems of identical elements was confined to the last chapter. Boltzmann strongly objected to th~ tendency of
Gibbs' bock: .II can understand statistics applied to a gasl; he said*, "but I see no point in applying it to asewing L'l8chine". Gibt;s' distrust of detailed ctomic·tUodcls was part of an attitude at cind shared by ~ny physicists towards the Gnd of the XIXth century: it was an essentially idealiEtic reaction against the mcchanist1cs school whieh had flourished durine th~ seccnd half of the century, and of which Kelvin· is the typical representative. Boltzmann rG'l.ilOined to the last
a staunch supporter of the mechanistic ideal; and in spite of the fact that this idaal has now boen recogn1ced·to be too narrow, ltd1d at the
. time give Boltzaann the right inspiration. Eis views cn statistical
mechanics \Tere pertectly claar and precise, although he lacked the
• aathcmatics, ve ncv have for their appropriate expression. y~t he had .. to face widespread acapticism and opposition and did not live to see
. the triumph of his' ideas.
~,t: ... :  ". ... y.
In fact, the objections raised agai'nst Boltzmann's. position
} ~ ..
•
• 0'
~
* This utterance was reported in conversation by Ehrentest who had been a pupil. of 3cltz~ann's.
.1 ..
, are only 1ntcrast1ng insofnr as they show the kind of' mlsundc:rstandins: of the sienif'ic:nce ot the statisticcl approach onc r:ight tall ihto. Lcschnudt 0bsGrvcs that if at a certain t1~~ all v:locitics (and magnetic fields) ot the system arc inycrtcd, the system will so to speak, rct:.."acc 1 ts steps: 11' thl.?::·efore, 1 ts ~voluticn went t.ruu less ·to more probable c,.::nf1gurat1ons, it will now exhit:it an 'opposite trend.
" , .
Zcrmolo, availing h1wsclf of Poincare's theorcm; pointed out th~t any
cC~"lf'iguratlon, hC:"love"r ir.1?robllble, 'nust bIZ expected to repeat 1 tsclf
wi th ar1.'itrary accuracy in the course of the time. Both assez ... tlons are, of course, p.::rfec,tly true: they can' be visuallzad by following the
path of the system ai:lcn·:; the ::ntanelcu(!nts of th:? various constellations. But they are tesl~2 the point: it is the c··~,arse d1stribu;t"on which has an irreversible evoluticn, and \I: have sean ths'i; such an irreversible tr~nd 1s p~rfectly compati~lc ~nth the reversibility of the mechansim by the lIm1x1ng~1 of the coans a distribution 1s brought about.
5. Systens wi til v;cak intGra ctl..9ps.
He may 'foUpw up the last argument more quantitatively. We
. introduce the? phase space of a single elcaent of the system, the "14  . spa cell in ',ihrcnte s tis terminology. The total phs se space vf a system
;
of If elern~nts, or ;, r apacev , is the direct product ct l:i identical f
spaces. Let us subdivide the},spoce into cells of arbitrarily small, but fin1 te size Wi; this defines a subdivision Of. r space into cells ccrresponding to the specific phases of tbe system. Thus, ccnsid~r the specif"ic phases [c·r which there are Ni elements in the fcal1 W 1 :
r !li Nt! 0 ifi
they occupy a cell in space of meaeure Wl9.J~ ••• , or f Wi •
To find the me asur c of the corresponding generic phase , we hove simply to multiply this by th~ number of ways in which the N clements con
be a.i:ranged 1n groups ot lrl~ Nc., ••••••• , 1.c. Ni / nll N~l •••• ; ,the measur~ at the g~neric phase 1s thus
,.
... )
Ni ;c Nl 0 ~ i N 1 1
•
To go on with cxplicit calculations, we must restrict ourselves' to a particular typa ct system: we assume that the intc~actions be
'·55<1.
tween its Gl~cents can be neglected; cxanples of _such systc~s.are the ideal gas and the pure radintion fiald. It 1s inportant to nGte that in order that the equilibriuu distribution ci the system be ergodic, it is essential that the elements do interact; otherwise, th~ n~bcra of elc!Ilcnts in the voricus energy shells of ~ space would cbviously remain unaltEred iq .the course of tiillC.
But the magnitude of the interuction 19 irrelevant and in
.,
suitable c1rcumstances can be regarded as negl1gible in comparison v1th
th~ proper energies of the elcoents. In this c.J,~c, we i::lBY assign a
. wclld~fined energy 8i to all elements in the ,M cell ""'1; llnd the total energy cf the system is an additive function of the, Ei'S:
E· E Ni E1 i
(lL)
If the total energy E 1s given, this is a conditicn 1QPosed on the distribution ct the elements ,among the cells£di• Another condition is,
or course?
(13)
t'le cay now readily evaluate the distribution N1 which, under the ccnditicns (~) (13) gives the constellation ot ma~imum QCasure (11). If the NilS ore large, the asy~ptct1c expression tor this distr1tut1on is
(14)
,\'1hcrc
(l~)
and the peraQeter ~ must be, determined in terms of E from the ccnditiOD (12.). In cvaluatioD; the IIsum ever states" Z it 1s 'pe!IDisfible totreat the jUcClls as infinitesimal, and to w~1te
J PH(p,q)
'l I: e : df,
(16)
'I
wh~re R(p,q) 1s the Hamiltonianof _a single eleoent, and the integration· is ext:;nded to the whole,.. spClce.
Let us nO\'1 (;valuate the measure n (1'111 , E
equilibrium. constellation, and ccupare it with, the
N~I , ••• ) of the ""E
measure of the
Cl1Cl~gy shell in r spccc in which it 1s contained. We have
lOB n 0111,.., NzIE, ••• ) ~ N log Z + P E, (17)
.£.0 •
so that the problea 1s reduced to the coaputation of log Z. To this
,cnd, we dccooposc the integration in~ space into on integration over 011 the values of thd energy and another over each cne~gy surface in yurn, using, of course, the invariant measure in accordnnce with
!;.f = df'E • dE • For the invariant mcasura of the energy surface E
~we "11'1 te .
fl (E) • eff" (E) ;
thus
+00
z = 1. e cr (E)~E dE
.00
The €xponcnt of the intGg~and hos a maximum for an energy Em given by ( d.'" ) • ~
dE f:)
Expanding it near the I!la xi mum , we ~ave, to a sufficient approximation,
. 1 ~r ~
C7" (E)  j3E =  (Em)  ~ED + 2 ( dE2 )Ill (EEr.l) ; :
the coefficient ~ (d2.(/' / dE2}m: is assumed to be negative. This defines a Gaussian distribution of the energy oruund the average value Eu. ~ho moan square fluctLliltion AEZ of tha energy is I (d2o / dE~) ml1 ,
which gives the phys1col.maon1ngof ,the, latter quantity .. To this approximation, we get for log Z
~og Z ~ log J.I (Em) pE + log
,. m .
~.logtJ ({ Eill ) ~Em '
it \tIe denote by [En.J an energy shell in ,... apace of thiCkne~S J ~:rr 4E2* around the energy surface Em. Inserting the last formula into (11) and noting that NEm~E, we see that I\. (J.IiIE, lIZ'E'''' •• } takes the i'orm
,.... ( {E l)N of the meaSUI"e ot the energy shell in r space conSisting
m _
of the shells (Em) in all theftspaces. In this sense, we :nay say
~hat ~ a.s.ymptot1callY; the equilibrium oonstellatioo is equivalent to
'Sl
the whole energy she 11 in r space. T,he distribution (l..4.), which coo
. .
pletely describes the equilibrium constellation, may therefore be used
'to compute, to a sufficient approximation, the ergodic averages of all additive phase functions for the special type of system he~e considered and, in particular, to define the thermodynamical functions pertaining to it. ;'Je shall comr back to this last point in a moment, after having discussed it for the most general isolated systems.
6. TherqQ9ynamics of isolated ·systems.
Isolated syste~s do not lend themselves very well to a discussion of thermal quantities, for the definition of temperature requires, from the phy.sical point otview,' the possibility of an exchange of energy tetween the system a~d its surroundings. Nevertheless, one can speak of the entropy of an isolated system and may thus ask for the atomistic interpretation of this quantity. The temperature will tllUS not receive any direct physical interpretation, cut will appear as a derived concept. We shall give later a more satisfactory treatment of this problem, but we may just as well take it up at this stage in order to present the general point ot view trom which it is :con s idered.
Let us first recall the peculiar way in which' macroscopic variab1es enter into the fundamental thermodynamics relations. Each nonthermal physical aspect is represented by a pair of variables (a,A) ,whose mutual relationship is characterized by saying that it the magni,tude of A is controlled by some agency external to the system, the ,work done on the system to change. a by da is A • da. In the atomistic

• treatment, we may take account of these macroscopic variatles (a,A)
: by introducing into the Hamiltonian of our system suitable terms of'
I •
'potential energy which .,'lill contain the parameter~ a. The associated
"variables A, called lI~orcesu in an extended sense, will then be de,tined as the statistical averages of the phase functions
(a/ a a) E_ (p,q; a) •
an isolated system, we must take the ergodic averages:
A = ~:lE
,

The thermal transformations are also formally represented by a pair of
,associated variables, the entropy and the temperature, but these must, of course, be treated separately and in a quite different way_
For an isolated system, the indepandent variables (i.c. those "lhich are controlled by external agencies) are the energy E and the parameters a. Th~ eljJtl~opy then plays the part of the ~ractGristic function, from which all other quantities are obtained by dlfferentia~lon and algebraic combination. If 9 denotes the absolute temperature,
one has
1 A
dS dEda,
9 9
(lS)
so that
•
! = (a) ,
8 J E a
A ..... s
  = ( ~) •
e ~ a E
(19)
2e shall always give t~e temperature, the dimension of an energy; the . entropy vill accordingly be dimensionless. Now, we can readily ind~cat; an expj;ession pertaining to the atomistic description of the sys
,teCl and which formally satisfies the relations (19) characteristic of the entropy. Let n represent the part of phase space ,ienclos~dl1 by the energy surface E, i.e. containing all the energy surfaces corre~
.
. sponding to energy values smaller than E. We may then take
S :: log t' ($1.)
(cO)
The veritlcction of the suItability o!this definition must _be limited to the second relation (19), since the first one can only ba,regarded as a formal definition of, the tempel'ature •. In order to ca1 :·culute the derivative as /ae at constant E, let· us introduce the
characteristic set function b (.0.) of the set·n, i.e. a function equal to unity for all phases of nand zero for all other phases; in the present case, we may express ,Ei (n) very simply as ,the I' step" function t (x) or the arguDlent x :: E . H (p, q ) :
. e, (x) :5 flo for x :} 0
~ for x .( 0
The derivative bt ~ (x) is the dl~tr1b~tlon i (x) , This allows US to write
·59·
J'A (n) j (; [E ,H (p,q; a)J df4
the integration extending to the whole phase space, '<!e further get
But clearly
fJ (E)
Therefore
and therefore
= ~ e
whence finally
A

9
In the speciol case of systems with weak lnteractions~ discussed in the preceding section, ",e may, of course, use the defin1 tion (~O) of the entropy; but we may just as well take tor this purpose a
_.
somewhat different definition, ~z.
8 • log n ( N1IE, NZIE, ••• ) , (n)
wh0ge explicit form is given by (11).. The "forcell A is now expressed as an ave~age over the most probable distribution (14)·
A ;:; J:L J 4J H di3H du
Z e « I'
the Hamiltonian H(p,q;a) now referring to a single element, and the integration teing extendedto the correspondingf space. From (~)
and (17) we derive "
(~) = ~
. a E a
•
•
. which gives the interpretation of the PQrameter ~ as the inverse of
60.
th~ sbsolute te~pe~ature, and
( =S) _ ,..H! ~h ~ E  ~' z ""Fa
·pH
e dLl = ·'"'A
, "
w:lich shous that the definition (21) of S is as acceptatle as (LO).
It is essential fpr fhis equivalence that the system ccnsidered has a It.rgc nu~ber of degrees of freedom; f0r the equation (17) 1s an asymptotic one. This aubiguity in the entropy definition for large systems
was discussed in dctoil by GibbS, and also from a more physlc~l peint
of vie\"I by Lorentz, who colled it the ilinsensibilityll of the entropy definition. The reDson for this lytitude in the choice of the function S is clear. He have seen already that the measure of !lO·~l1E' rf~IE •• ) is practiccilly the same as ,that of an energy shell in r space; and the latter is not Significantly dlffe~ent from the measure of the whole domain1"l when the number l'T of elements is very large. To give a
simple exampla, consider an ideal monoatomic gas, vh~e Hamiltonian consists of the kinetic energy of translation,~ 1':1 Pi L (m being
. ~m i4
the mass or an atoQ) and of a potential energy expressing that the system is confined to a volume V. In cocputingf<n) the integ~'aticns
, over the posi tioD and .nomen tua coerdane tes can be performed. sepcrately, the foraer gives a factor ~{, the latter the volume at a 3trdi~ensional sph~re of radius J2m E, which is proportional to E(~!/Z). The measure jA{E) A E of an energy shells differs from fJ(.fi) only by the replace
. ment ot E{3Y/Z) .. ,.. ., 3(Nl) A E which for very large N 1s
by (3N/l) E z ~, ..  
quite insensible.
Our tuo definitions of the entropy have a common teatta'e:
they both express the entropy as a logarithm of the measure of a domain of phase space, i.e. ,of th~ relative probability of occurrence, in
the course of t1~e, of the phases contained in this domein. The logarithmic character of the link l:etween entropy and probability 1s easy to understand: if we unite two independent isolated systems into a single system, their entropies should be added, the probal:ilities of the ir configura tions  mul tlplied. The de tin1 tion (~l), usually adopted . for systems \nth weak lnteructions, corresponds to a more detailed discrimination of the COi1r1guratlons than the general definition (20).
'.
61
The law o£ increase ot entropy appear~ as a statistical
law. It (by removing constraints) we allow the isolated system to pass trom an initiDl state ot equilibrium to the state of equilibrium co~responding to the new conditions, the va»ues ot the entropy in the tinal state will necessarily be larger than the initial one; but the behaviour/ of the statistical function log • {lOll' NZ_ ••• ) during the transition will, as already stated, exhibit continual fluctuations (even afte~ the new state of equilibrium is reached).
The concept of entropy, however, essentially refers to the state
of equilibrium under neglect at such tluctuationa. The statistical interpretation thus indicates the limit of validity at the entropy concept: if the fluctuations form a prominent feature ot the phenomenon, as in Brownian motion, the thermodynamical concepts become meaningless. The second law 1s not "violatedll, it ceases to be applicable. On the other hand, the statistical treatment is still perfectly adapted to this kind of phenomena; e.g. it can be used to evaluate the average amplit~de of the observed fluctuations. In this sense, the scope of statistical mechanics is wider than. that at phenomenological thermodynamics.

ee
..
III .. STAXlbTICS OF CLObED SY&Tr;HS
1. Syttem of given temperatur,e;
To ensure that a system S has a def1nite temperature,
we iJUSt il:lagine it in ., thermal ccntact« lY'i th 0. II thermostatll, i.e. some other system T vhose heat capacity is so large that the exchange of heat \"lith the system S does not sensibly alter its temperature. The two systems S and T, freely exchanging energy with each other, have then a 'common temperau~c, which 1s. entirely fixed by the thermostat. From the atomistic point of view, we have a total
,) h
·systcm~ consisting of t e system S under investigation end the
thermostat T, such that the number of 'degrees of freedom of S is much smaller than that of T. Uh11e the energy of beth S and T continually varies, we may assume that the total energy of J remains constant, i.e. that J 1s iso1ated. The problem is thus to derive from the statistical distribution of the isolated system ~ that of a small part S of t$ •
Fer this purpose, we consider the phasesl8ces of S and T, '. with measures df s·' df' 'i and the product space of ~ wi~ measure
df = df Sdf'T Moreover, although the systems Sand Tare es sentinlly interacting, we may assume that the magnitude of the interae'tion is negligible, so that the total energy E is composed addlt1vely of the energies ES' EX of the systems Sand X •. We again 1n~ troduce 1n the S and Xspaces and in the product spaces the domains P.s,Ox,.n. enclosed by the energy surfaces EX' ES' E. NO\'I we
a~c interested in the averagas of phase functions pertaining to the system S, i.a. depending only on the phase coordinates Ps• Ue
may write down these averages iCi.UD.edlately as ergodic averages in the totol J space; ~e shall then see~ to transform these expressions so as to eliminate the irrelevant phase coordinates PT and
,
.
;:..
, r
63
the thermostat will be represented in the tinal result by only one

parameter, which plays the part of the temperature.
For the transformation we have in view, it will be convenient to write the ergodic average in the form
liE = L 2 l . r.:(p) df (1)
J . f1 (E) dE J n
For a phase function repS) belonging to S we may carry out the integration with respect to the variables PT; for ea.ch phase Ps' to which corresponds an energy ES = Hs(~), and accordingly an energy ET = E  lis, _ this gives a factor f' T [:n T (E,  Hs>1 • The integration over Ps extends over the dbmain enclosed by the surfaces corresponding to the large~t value of ES' viz. E; but we may actually extend it ove~ the whole Sspace since tor larger values of ES' the tactor f T vanishes. Thus,
fi(Ps> r = f reps) /4 T [ar (E  Hs» df s •
'.
The derivation with respect to E transforms P T(nT) into the invariant measure of the limiting energy surracejU T (E  US>: and we
get r4'om (1)
(2)
tIe have now expressed the ergodic average ot f(PS> as an average over the phase space of the system S alone, each element of this space being affected with the statisticalwe1ght/J T(EHSJ.~(E) i since this weight fac~or only depends on the energy HS' one may say that it affects the energy surfaces or energy shells in S~space.
As a result ot the contact with the thermostat, the energy ES of the system S is not fixed, but has the equilibrium distribution given
.!:y f'T(E  ES)/f" (E): we must therefore take all ergodic averages fiE and add them up with the weights corresponding to the distri
but!on of ES. .
2. gpnon~c!l distribution.
Our next task is to find a simpler expression for the energy. distribution in S  space. This means ~ssent1ally that we want to·
•
64
transform the factor f'T(ET), for the factor /'A (E? is eas~ly evaluated in terms of it; in fact, by an argument similar to the above,
we find
(3)
the integration extending over all values of the energy ES.
I .
Now, we note that the structure of the thermostat 1s en
tirely arbitrary, apart from the requirement that it be a ve~y large system. \re avail ourselves of this latitude to obtain an asymptotic expression s=r T(ET): to this end, we ~ssume the thermostat to consist ot ~ large number ot par~s TI, T2, ••• whose interacti~ns we· may neglect. By making repeated USe of a fo~mula of the type (3)
we get
f'4T(ET) = f dET J dET • ···fdF.r }J T (:~T,,') /.A T (E1' ). 7'T (~l .. E~f ~~T
1 2 ml 1 _ Z. 2 n 1 ~Z
The formal analogy of this equation with the 'ilaw of compcsitlonn
. .
of probabilities suggested to Khinchin an elegant application to
our problem of the Ucentral limit theoremll of probability •
. Consider n stochastic variatles ~l' xZ, ••• Xn \lith independent distributions ul(X1)dx1, Uz(xZ)dx2, ••• The distri~ bution of the sum x = Xl + Xz + xn .oteys the law of composition:
u(x)dx = J dx1! ~2. •• J ~l 'uf(x1) ~2(XG!) ••• un (xxlx2···~_i) (4)
Now, prcvided that th~ distribution laws ~(Xi) satisfy certain cind1tions the most important being that the mean fluctuations of the variable Xi are finite, the distr1bution law for x has the
asymptotic form (tor ns ee ) "'l.X~)Z
u(x)dx~ 1 e  ~~ (;)
 .f27rB
,,,here ,x = ~ ',~ is the sum of the average values Xi = J x:L~ (~)~, :.
•
(,5
i.e. the average value of x, and B = "i: bn 1s the su~ of the mean
quadratic fluctuaticns n
1:i,=/ (~~i~1)2 Ui(~) Wei
We cannot, however, immadiately identify fiT (EX }dEX 1 1 i with a distritutif~ l~w ~ (~)dxi because the measures fJ Xi (EX)'
1 though finite, are not necessarily bcunded: they generally increase
as some power of the energy. Th1s obstacle 1s, ot course, easily overcome by putt1ng
() 1 PEXi u )
~ EX! = Z1(P) e r t (ET
. 1 1 .
(6)
l'lhere
Secures the ~crmalization, and 1s an arbitrary positive parameter. In fa ct, the Uti S so d~flned aloe bounded and satisfy the law of COlaposition (4); it can be verified in detail* that they satisfy all the ccndltians to~ the validity of the asymptct1c formula (5).
The law of distribution U(ET) derived from the set (6) by the law of composition (4) has exactly the same form as each
of the ui' s, the norma li za t.1on fa ctor Z (P) be 1ng the product of the Zi'S. We readily get
ET =
and
d log" Z dP d2 log Z
dpZ
(8)
(9)
It 1s easily seen.* that equat10n (8) uniquely determines ~ 1n
terms at Et: this makes it poss1ble to fix the parameter J1 by relating it to the given average energy of the thermostat.. We shall henceforth assume that P has the value resulting from this equation (8): it is this parameter which will "representu the thermostat in the final result, and we shall see that it 1s simply related to the tempe rature.
________ . __ !;'?l9:I.!S ... !2!: u(~) its asymptot1c torm (6) we get for
• Vc1r Klh!nch1n, p., ~6
*0' Voir Khinch1nj, p&. 77
•
£5
U (E ) _ Zx{P) r: T T  '2iTB:
v "X
(10)
In applying this for~ula to the argument EX = E  ES' we note that (ET ET)G = (Es /ES}4:!. Now, excepting the unlikely cases when ES happens to be widely different from its average Es, ~he quadratic fluctuations cf the system S will be of a much smaller order ot mag~: nitude than those of the thermostat whose average is given by Sr:
we shall accordingly neglect them altogether and write
ZT(fl) P(EES)
f'T{E  ES) 7Z1TBi e
For the factor fA (E) we may write down a completely similar expression since according to (3) the systems Sand T are linked together by the same law of compos1t~an as the various parts of T: we may the~efore treat the total system S as we have just treated the thermostat. Thus,
and
while B differs from ~ only b~ the contribution from the small system S which can be neglected. For the energy distribution of the system S we therefore get the very simple expression
(ll)
in which the thermostat only appears through the parameter
p.
In its most general aspect, the property expressed
by formula (11) may be stated as follows: any small*:":part of ;ItneedhSrdlybeemphasized that the "small" part' can be any ma~· croBcopic system; it' is smal1 only with respect to its surroundings which constitute the,lItherm.ostatu• On the other hand, the small part could also ba just an atom of an ideal gas: Boltzmann's theorem then "gives an 1nd~pendent derivation of the theory of systems of weakly coupled 1den~ical elements wb1¢h we have studied.
•
61
n large system, whose interaction with the system is weak, is distributed in energy accord1llg to the law (11). In this feum, the theorem is due to Boltzmann" who recognized its :fWldamental importance for the statistics of nonisolated systems. The law of distribution (11), under the name at canonical distribution law, was extensively st~dif~ by Gibbs; the insufficient emphasis put by Gibbs en BcltZIJaIlll'S theorem is no doubt responsible for the widespread confusion about the physicol significance of the canonical distribution. The above derivation, due to Khinchin, has the merit ot elucidating the deeplYing relation of Boltzmann's theorem to the general principles of statl$tlcs.
.68
3, Energy distribution.
The canonical distribution law has important consequences for the average value of the energy. In the first place, the study of this quantity will allow us to establish a very simple relation between the thermostat parameter p and the absolute temperature as usually defined. Further, we shall derive a general property known as lIequ1partit1on of the energylf. Finally, it is also possible, at least for a certain class of systems, to derivG the explicit law of distribution of the energy around the mean and, in particular, the mean quadratic fluctuation of the energy.
The canonical average orthe energy of our system S 1s given
by:
If, in particular, the system S is an ideal monoatomic gas, we have
"." . f' (ES) ,....,; ES ~.~l
, whence
and
dEs
3N
,101 ~ T
,"
(12)
The thermodynamical value of the energy of such a gas is ?
~ R T per mole, where R 1s the gas constant and T the absolute ~
Z • 3
tempe~ature in the Kelvin scale. This may be WTitten.Z N e, if N
is the number of atoms per mole and a = kT, with k = !VN, is the measure of the absolute temperature in an energy scale. The conversion with the statistical formula fixes the relation ot the thermostat parameter P with the absolute temperature:
13 = .L (13)
9
T~~ simple formula (12) is a special case ot a more general
~
propertyo It we can isolate 10 the Hamiltonian ot the system any
group ot terms H' homogeneous and quadrat1c in· some ot the pIS, i.e.
69
_ 1 k ~ H'  2: L Pi dPi
, 1=1
we may perform the calculation of the canonical average HI explicitly. In fact, putting H = HI, we may write:
iI· = ~J dr u e PH" f U' eIlH' d 1" ,
i
where dp = df' dfu with djJ' = relevant Pits we have
( >HI a ...... • p ~ a·ll) dp =
J op
whence
Now, for any of the
Ht . k 1  k ' A
 ....... .... .Au.
2. ~ 2.
(14)
The same argument applies to the case that the quadratic homogeneous part of the Hamiltonian belongs to the ,potential energy, a.e. has this property with respect to some of the q's.
The general formula (14) expresses the law of eguipartition of the enertl: for a system of temperature e " every degree of free~_ dom for which t.b.e kinetic or potential parto! Haml tenian is, homogeneous .. and:· , quadratic in the corresponding canonical variable contricutes ~ e to the average energy of the systelD). In particular, every degree of freedom of harmonic oscillation contributes ato the
average energy.
If the system Scan itselt be analysed into a large number or component parts with weak interactions, e.g. if it is an ideal gas, we can apply to the mBasure of its own ~;energy surfaces fs(Es) the asymptotic formula that we ha9 derived 1vr the thermostat. The law of distribution of the values Es of the energy then takes the
c1mple form: . (EsEs)Z
WeE) LI (E )eT(EEsl_ afP! ePEs 2Bs,',. Z (r.)
s = t S S }I (E) ...(Z1rBs ''T t'
,
f!(EEs)  e
(EsEsf
2 ~ , 2Bt'
. t.(p {_ {EsE's )._ I •. 2'D'B e ss, 1 e
2Br ~Z ( ~) . ..J2'1tFlE
70
where BS 1s defined by 1
=
We thus obtain a Gaussian distribution of the energy with a mean quadratic fluctuation Bs * which 1s always smaller than Bs: the exchange of energy with the thermostat thus results in a negative correlation of the/energy distribution between the twopa;·ts of the tota~ sys te.o.
The value of the mean quadratio fluctua~1on Bg is aas11y
computed. One has
•
d log Z dp
,
~_+l~Z ~s . Z dp2 ,
dES

dll
(15)
This formula exhibits a .f'undamontal !.eature COCJlllOll to all
. dE
fluctuation phenomena: while the product 9 ~ 1s of macroscOpic
order of magnitude, there is an extra factor e which contains the
essentially atomistic constant k. In fact, fluctuation phenomena are outside the scope of macroscopic ,physics: they are essentially linked to the atomic theory of matter and radiation. Accordingly, the observation of their macroscopic effects, such as Brownian wotion, light scatterin~, etc., affords'a proof of the consistency of the atomic picture and a possibility of determination of atomic dimensions.
4. Thermodynamics of closed systems.
We shall now follow a line of argument parallel to that developed for isolated systems. We shall look for an atomistic inter;:; pretation of the characteristic function corresponding to the choice of the temperature and the external parameters as independent variables: this is the ~ energz of the system, defined as *
 ..   . .. ..  .. ,. .. .,.. .  F = B  e s . (16)
* We may now drop the index
71
In fact, we find
i.e.
dF = s de:!: Ada,
S =  ( :!~ , A = (~)9
(17)
(18)
From (16) and (17), we derive, by making use of the general t.L'ans.'; formation forlIlLllaJ:·
e aF  r. aF
is t'~,
...
the Helmholtz relation between mean energy and free energy
 aF:'!\,
E = F  9  = ~ (PF)
ae ar>
Now, a direct calculation gives for the canonical average
of the force, (JHJ c)a)
A = _! alaS' z .
!3 a* '
this formula, together with the previous one
E = _ alog Z QP
shows by comparison with the second equation (18) and with (19)., res~ctively, that we must take fO,r F the, expression
..
F =  & log Z,
(20)
except for an arbitrary linear function of 9 (without physical meaning).Formula (20) completes the atomistic interpretation of the thermal quantities pertaining to closed systems, viz. temperature and free energy. The entropynow appears as a derived quantity:
S = log Z + P E
(21)
It is not surprising, that this expression for the entropy is fcrmally similar to that we have obtained for the systems of identicalelcments With weak interactions: for, as already noted (see footnote at the end of paragraph 2), the theory of such systems can be regarded as' a special applica·tlon of that of closed, ."
:!Z~~~!!::!. __ !i£>~.Y~':.lJES_':;E.i>.P~_2!_Iormula (21) is much wider : it
1." The system S then represents a single element, and the expression (21) accordingly,gives the entropy per element.
12.
.
applies to any kind of system without restriction. If the system
is sufficiently large to allow the usc of the asymptotic expression for the measure at its energy surfaces, we cay compare the value'(2l)
,
of the entropy with the logarithm of the measure at the surface cor
responding to the ~ energy E, viz.
. I· log ft (E ) = log ! z ( P ) e Pi} = s  log (..(2.1iB) LJZlTB
S is log measure of energy shell as before.
With the help of (21), we easily verify the wellknown thermodynamic property, on which the phenomenological definition
of the entropy rests, that the i~verse temperature p 1s an integrating factor of infinitesimal quantity ofheat supplied to the system in a quasi.static transformation. This quantity of heat is
'Q :: dI ... Ada
whence
13 ~Q = d (f~)  E dP  p Ada
d (pE) + d log Z d P + dlog Z da
=
.~~ • > aa
= d (pi) + log Z) = d S Finally, let us introduce with Gibbs the concept of pro, bability exponJLn~ which will give us notonly a compact notation for canonical averages, but also exhibit a new aspect of our entropy definition. The canonical average of any phase .function
f::= ~ J f (p) e"';flH df
may be written as
i =f (p) e ~ (P) d f
with
~(p;e , a) = ~ [F(S, a)  H (p;a)]
, (lZ) (23)
The function is called 'by Gibbs the "probability exponentU, since 'Y)
~ (P; 9, a) .~ « (24)
represents the density of probability in phase space for the canonical distribution. Now, the entropy, according to (16), is just minus
73
the canonical average of the probability exponent:
.
s = :;.;
this may be written, with the notation (G4)
S =  f n log n d tt
Again we find here a formal analogy with Boltzmann's Hfunction for syst~ of weakly coupled elements. The expression (25) may be applied to quite general systems: it does not then refer however, to any actual distribution of physical elements, but to a purely fictitious distribution or IIcop.i,.esl1 of the system un . der investigation.
The statistical foundation or thermodynamics just outlined is not yet logically complete. The easic differential equation (17) refer·s essentially to guasist..!ttic transformations. We still have
to investigate the behaviour of the entropy when the system is subjected to irreversible transformations. But beiore thiS, we have a more immediate task. In the preceding discussion, all macroscopic quantities have be~n as a matter of course represented by canonical averages: this procedure, however, implies the assumption that the quasistatic transformations do not disturb the canonical for~ of the equilibrium distribution. Quasistatic transformations are or two dit~inct types: isothermal or adiabatic. For isothermal· trans .
. formations, the assumption just mentioned is obviously justified since the system remains in contact with the same thermostat during the transformation. But the case ot adiabatic transformations requires a closer investigation. In fact, in such a transformation, the contact with the thermostat 1s initially severed, and at the end the temperature of the system has varied, i.e. it is in equilibrium with another thermostat: it 1s then far from obvious that the distribution, while both external parameters and temperature varied, nevertheless remained canonical. This is what we shall
now proceed to show.
5. Permanence C?t...9Al!QR:1..9A~tribution in quasistatic adiabatic transformations •
........ .....~ ......... ~ ... ~..................  ........
An adiabatic transformation can be analysed as a succes
74
sien of steps in which" the system being isolated, some external parameter (or set of parameters) g is varied suddenly by a very small amount da, after which the system is left to itself £er a certain time. A second sudden variation of Q. is then applied and the process is ,repeated with suitably chosen intervals between the
. successive variations so as to approximate any continuous variation aCt) or the param~er~. In order for the transformation to be
'quaSistatic, we must, after each small variation of A, allow sufficient time to elapse for the system to.,reach the state of equilibrium corresponding to the new value of the parameter: a quasistatic adiabatic transformation 'corresponds in' this sense to an infinitely SLow variation of~. •
For our present .purpose, it 1s of course sufficient to consider a single step, leading from ao to ~ = ao + da, and to show that if the initial distribution was canonical, with a certain temperature (or, in Gibbs' terminology, "modulus") 90, the final distribution is again canonical, with a slightly different modulus. The variation of ~nas the effect of deforming the energy surfaces so that a distribution which was originally e'rgodic ceases to have this property when referred to the. new energy surfaces. Eventually however, it .will becomeergodic on these new surfaces, and the new phase density wi~ be given by the ergodic average of the old one, taken an the new energy surfaces. The initial phase density was
gi~en by e::,"'(80,ao; p q) ; ~he new one is accord1..'lgly .
.,..,(80, ao;p q I .
0'' E(a)
1;1e, want to show that this quantity can again be written in the form'or a canonical distribution of suitable modulus B.
To this end, let us f1rst exp!lnd the dlstribut:1,on density
'J (9o, 80; p q) .
e around the new value a of the parameter: .
?J (90, ao; p q) ." (90, 8; pq) C' da . _H )
e = e 1  (A  
. ~ ~
1)(90, al pq)
and accordingly, sinc~ the factor e 1s constant an the .
new energy surface E(a)
15
?'} (60,ao; pq) I .,,(eo, 8; pq) [ d J
e E (a) = e 1  ...! .. (A  A I )
90 E(a)
Now, for systems with 8 large number of degrees of freedom , the canonical jistributian has a sharp maximum for the @verage I and
we may use
A = AL.
E
•
, '
The difference A ' AI_ will thus involve, besides a term linear in E  E a term conta¥ning
2 2 2
(E  E) (E  E) ; (E  E)  B
I
this term will be negligible in comparison with the first one, and we may therefore "rri te
~ (80,' 80; pq) I ~  (eo,a;pq) [ d~ '~AIE  ~
e  E ( )  e 1 +  (EE)
a 80 ~ E'
It 1s now clear that the correction term in E  E can be compensated by a suitable change of modulus; ,for if we pass from eo to 9 = 90 + de we have, using Helmhcltz's formula,
. ? (Oo,a; pq) '. "(9,a ;pq) [1  2!z. (EI»)
e = e 90
lie conclude .that the final distribution 1s indeed the canonical one of modulus Q, provided that the variation de of modulus 1s related to the variation da by the equation
'Jo clAi
' ee Ed' 0
  a =
90 (:)E
'by. A and 90 by.,S, and not1ng, at constant §
~.=~ '/ ~
oE De as'
ReplaCing Al_ E
the last equation may be written 1 ~ .cle _ 8 be
a~ da = 0 . as
76
It is readily vcrif!ed * that this equation just expresses the constancy of the entropy, dS = 0, during the transformation. 6. Irreversitle transformations.
The last step in establishing the statistical foundation of thermodynamics consists in verifying that our interpretation
of the entropy mynus as the average probability exponent of the canonical distribution satls~ies the part of the second law reterring to irroversible transto;"mat1ons. . \'10· must show. :that
:tn an:' ~.'i'I"rt:vcrsi blc passage of cur system from one oqu1 l.ihr.1um~:sta tc ,. to  another, tlm entropy ,.. or.: .t~ ri~l state
of the system is larger (or at ?ny ratenot·smaller) than that of the initial state; if we formUlate the law in terms of average probability exponents, th~ direction of the in~quality 1s reversed. It must be observed th~t a probability ·expanentwhich would corre'spond to an equilibrium state under certain conditiQl?s represents an arbitrary dens:!,ty distribution if these conditions are altered. \'ie shull teus hove to introduce, following Gibbs, arbitrary probability expcnents (P) and study the behaviour ot
their average values. _
UJ=fweWd~
under the various transformations. ;satistylng the abovementioned condition. All suci~ transformations reduce to two types: (a) the adiabatic transformations, during which the system is ~ermally isolated, (b) thG heat exchange, in which two initially separated systems a:e brought into thermal contact and separated again. We shall discuss these two types in succession, closely tollowing Gibts' elegant treatment of the proble ••
(a) 1illj,al::2..t;'g • .1r1!!1sformations. During an adiabatic transformation, the distribution changes independently in each
* In tact, from (18) it tollows that _.~ = J.!
de aa
and from (19) tl"..at' 1 E _ ctF _ QS
...... ,..~~
e 9 .  ae2 as
77
energy shell: we shall thus compare distributions with the same total number of' "copicsil of' the system in the same energy shell, and prove that thc distributions with probability exponent constant in each energy shell have a smaller average probability exponent than the others." We call this statement the first lemma of Gibbs. The assumption 1s
for each energy shell (E). For the average ot ~o over a shell we
\Ai { e \U d, A.I ;. J rt. eW dLJot
o J (E) I . (E) 0 I 
and therefore
'\tIe must now' prove that
~ ~ ~ ;. J (w wo) eW df4 ),0
Gibbs uses here a ~eneral procedure a,ppllcable to any of' the form  f xe df' ~ O. He refers this inequality perty of the exponential function:
eX  1  Xq 0 for any x.
To this end, he uses the normaliZation conditions to show that the appropriate terms eX  1 associated to x have the aver
_ age O. In the preseht instance x = W ~; f'rcm the normaliza .
tion condi t10ns J W f .:Ai
e df' = e 0 df' = 1
inequality to a pr~
.
it indeed follows that
 r W. w ]
JeWle 0 1 »:>
expression Jew  G.'o) e W df', we immediately quantity 1s always "'0.
'.
•
Adding this to the conclude that this
Let US now copsider the successive steps ~ of the adiabatic variation of the parameter ~ from ao to~. Let VJl, ~ •••• ,~ be the successive probability exponents. It the transformation is
qua 91s t.:~ tic 1 ~ 'he di ffel"encfw:;J
78
.  ~ ,
quasistatic, the d1ttcrcncwS' ~::;:lk~';; are of the second order in
,ds., the difference iLJ1 ~ after a finite variQtion at .! 1s 1nf"ini~ .. _tesimal: ~he entropy in this case remains constant. If, on the other hand, the transformation 1s irreversible, the successive differences will 10 general be of order d~, and for afinite variation of.! there will be, by Gibbs' lemma, a finite increase of the entropy.
I .
(b) Heat exchange. To discuss the case of heat exchange· we have to consider a system composed of two distinct parts 81, S2; we accordingly d1stinguish the two phase spaces 01' these parts and the total phase space which is ~heir direct product. Between the respective elements of ~easure we have the relation
d}l i .' djJ2 = ~jJ
The distribut1o.11 eW def1nes the distr1butions in the 81, and 82 _
spaces as aWl =JliJ dl'2 e 012 = J eW dl'l ;
these partial distributions are, of course, normalized to unity; Je WI dfl = f e Wz. dflz. = few dl' = I
Gibbs' second lemma states that
  
Gia~+W2
the equality holding only if. the two distributions iAi,~, are entirely independent, so that w=~(P) +WZ(P2)
To prove the lemma, note that
::. J  Wl J  W
~ = WJ. e df 1 = wI e d}J'
and therefore
The normalization conditions give
f  u:, + lil..W
, e W (e J: ~ 1) dp = 0
whence the announced.1nequaI1ty tallows. Moreover, the equality is seen to nold onlY it' ~ ~  tuz = 0, as stated.
With the help of this lemma, the discussion of the thermal
"
ccnt~ct of two systems is im!.HJdiate. Before the contact, we have independent distributions with exponents ~l'~2 and the exponent
of the total distribution is W = W1 + l~ ". j~S a result of the thermal contact. (during which thO' total system is isolated) the exponent changes from ~ to~, and by the first lemma
I
~ =,
W ,?W .
llot/, we had on the. one hand w=iAi+CU~, " and on the other, by the second lemma
 = :.
Wi ~U,)i+WI
• 2 '
WI]! (Ai 2 be ing the exponents of proba bill t,y of the. two systems in their final states. Therefore
Wl + G)l ~iJl + m2·
(26)
i.e. the heat exchange has caused the sum of the entropies of the two systems to incrGose.
It is interesting to analyse somewhat more closely the role of the temperature in the phenomena of heat exchange. The , basis for such an analysis is supplied by a third lemma of Gibbs:
If e i~ a positive parameter, the quantity ~ (E I e) 1s mtn1mum for'the canonic&l distribution of modulus e. Fer the latter dis
tribution, the 'quantity in question is just Fie; the lemma thus ex· presses a DiniBllm propeFtyof tne fr~e energy of a system of given temperature, which is parallel to t~e maximum property of the entropy of an isolated system. He have to prove
j'(W + E/B) ,eW • df  Fie ~ 0, i.G. !<Wrt:!) eWdf~' having the probability expdnent of tho canonical distribution of modulus e. By the usual procedure, \/e derive from the norrJBliz8 tion conditions
whence the preceding ipequality follows •.
Suppose one of the two systems just considered is a il.thcrcostat" of temperature Sol. Bes1des the inequality (26) we may:, now \Ir1te, by the third lemraa,
86
since the initial state was described by a cononical distribution  of modulus QZ. Therefore,
I . W~  Wz. ~ EZ~!~ (.m
or in terms of thermal quantities
•
,
i.c. tb4 entropy increase of a sfstem brought in contact with a thermost,at of temperature Q2 1s at least '1/92 , where Q 1s the quantity of heat passing from the thermostat to the system. Using
EZ  B'2 = Ei  11
we can also write the inequality (21) in the form W +!l ~ w' + .!'
1 t:fZ V 1 6z
, = '
showing how the quantity ~ + (~ / 62) pertaining to the system 81
decreases after thermal contact with the thermostat, until, by repeated or prolonged contact, the canonical distribution of;mbdulus 92 is reached, .and with it the minimum value of' the quantity in questicn, which is then the free energy of the system.
\'Ie may finally consider a system So undergoing a cycle of transforms tions in which it come s succes s1 vely in' contact wi th various thermostats of temperature.61, 92 •••• Assuming So to be 1nitially at temperature 90, we have
Wo + GJ1 + mZ + •••• ~u.)b +Wi + l:52 + ••••
(1 = O~ 1, 2, .).
whence
.J ~._~ ....
(ZS)
the term pertaining to'the system So has disappeared since E'o=Eo
81
'uine t o th,:··cjlclic charnc tcr cf the transformation. The inequality (28) coincides with an importcnt .theorem proved by Clausius
and used by him for establish1ng on a phenomenological basis the concept or entropy. If there are only two ,thermostats, we have the kind at idealized engines considered by Ca~not •. Calling Ql,QZ.the algebraic quantities of heat supplied by the thermostats during
a cycle, inequality (28) takes the torm .
~ + ~ £.0 OJ. '£
If the heat transfers are irreversible and not accompanied by any mechanical work, ~ •. g,. it heat .is conducted or radiated through the system So' we have Ql = Qa,. an~ the inequality .C1.erely expresses the tact that heat pas~es irreversibly tram the hot to· the cold thermostat. It an amount of mechanical work W = Q1 + Q,
1s supplied by the engine (~ > 92; Ql >'0), we f'1nd tor the et ... f1c1ency the well known expression
':! c ~~a , Ql ~ ,
the equality corresponding to Carnotrs.~eve~s1ble erig1ne~
..
82
IV.  ST .... TISTICS .QE ~ SYSTEl.fS.
I .
1. Phase functions and averages tor open systecs.
Our last object of investigation is the statistical description of systems consisting ot a variable number ot elements. Since the main appl1cction ot this theory is the study of chemical reactions, we shall speak of systems of "moleculesli• We start from the ft' spaces of the single molecule~, from which we construct by d1rect product the rspaces corresponding to any g1 ven numbers ot molecules of the· var10us species occuring in the system considared. ·In the following, We shall only treat the case of a mixture of two distinct constituents, which is sufficiently typ1cal. We denote by ~1~2 the set of cocrdinates·of the ~space corresPonding to Nl~molecules
of the first species and NZ molecules of the other.
The det1nitlo~ ot phase flUlctions requires some care.
In general, a physical quantity pertaining to a syste~ of N ~lecules orthe same s~ecies may be expressed in terms of other quantities attached either to a single molecule or to a pair of mole.cules or, more· generally, to a cluster of any number ot molecules. It cnn thus be written in the form
f(PN) =. t ~(l) (Pi) + t t f'2~P1 ,Pk) + ••• (1)
1=1 1=1 k=1 ...
vhere P1 denotes a.phase in the ~space o~ the ith molecule. The extension of 'this definition to the case of a mixture of molecules is immediate: it gives a welldefine~ meaning to the notation
f(~lN2~·
In order to compute the statistical average of such a phase function for an open system, we follow the same line of argu
•
ment as for closed systems: we consider our system in conjunct1cn
with the requisite number of very large reservoirs of molecules at. the differe.rit specd.es , so that the total system·' may be regarded as/'
...... '~ .... 
.••• ~ ~~. .+ .......... 
__ ~'4~._ ....... ~ • _ ......
83M
c10scJ.. t'or tho' closl.idsystcm, which we OSSUrle to have a definite
~~~pvr~turc, we @ay write down the canonical aver?geof the phase llulcti0n. There then reaoins the task of eliminating the variacles pertaining to the reservoirs. The exchange of molecules between the system S containing a _mixture of different species and any reservoir R containing one definite species must be ~magined to take place through a suitable/semipermeable membrane: the physical existence of such ~eabranas,. howe7er, is not relevant to the argument; it suffices that tho concept of sem1perceablo menbrane be logically consistent with the basic assucptions of the atom1cth~ory of ~tter.
The phase functions referred to the closed systC!!l
S + RI + HZ are themselves' statistical averages over all possible distributions of the "molecules between the system S and the appropriate .. reservoir. With respect to each species of molecules, the systen S can be regarded as a small part of th~ corresponding large reservoir; the law of dlstrj.cution of colecules in S therefore takes the asynptotic forQ ofa ~o1ssan distribution *
FP N
lI(N) = e
N!
N being the average number present in the system S.
(z)
Thus, the phase
function whose ~Q~onical average 1s needed is
L; lI(Nl) W(NZ) t (FN1~) , (3)
!ll,liZ " 2
with the definition (1) of r(PN~2). Here, f(PNINZ) depends only on the variables of the system S. Ue have also to consider the total Ha~i~tonian of the closed system S + Rl + R~, which we may take to be the sum of the Ham11toni"ans of the" t~ee parts:
_ H + ERl + RaZ. For any distribution of the mol~cules, characterized by the numbers N1, Nl, we have, more explicitly
H(P.~l.lf2) + lit! (PNi) + HR.a(PNl";) (4)
where PN:" I denotes" the phase of the reservoir Ri containing. Ni I
....
colecules. If J'i represents the total number of molecules of "
species 1, we have Ni:: Hi  Ni" so that the phase function ofthe core general type (4)~ is also specified by the nunber~ N1,NZ
1
~
* He eight even take a Gaussian distribution but the choice of the Poisson distribution is somewhat core general and tits in better W1th
__ ",,=.r:"
the thermodynacical formulae.
84
~r ~ol~culas in the systcn S.
The statistical weight of th~ phnse PNll1Z of the system S 1s therefore prcpcrtiannl to
. PH(PlJl,NZ) I ~ [ HIi' (PN;') +",Ha2 'PN2) 1 I
df'Nl ,N~ e . . c '. df HI d.,u NZ '
i.e.
where Zni{Nl) represents the !lsuro·over states" for the reservoir
Ri containing Ni nolccules. Denoting the c~rrespond1ng sue over states tor the system by Z{N1,NZ), we finally get for the canonical.
. ~verage of any phase tunc:::tion pertaining to the syst7m S the expression
' .. !'
c,
In this formula, the parameters defining the 'external conditions under which the system is considered are, besides' the tempe~'ature and the nonthermal parameters of type A, the average numbers NI, NZ of molecules in the system (which,.however, are not neces
• • •
sarily all independent). _ .
2. Chemical potentials.
'" r
The problem now confronting us is to find some Simple asymptotic expression tor the dependence of the function Za.t. ~N:i.)' representing the res~rvoirs. in formula (5), on the numbers Ni of molecules' contained in them. ' S1n~e the total num
be~sJtl can be made arbitrarily large, we are actually 1ntereste~ only in small relative' deviations ., :~ .. ~( Ni  Hi.> / N! of the
. numbers N:i. from their average values !i1 = . .Jf 1;'''''' Ni Even the absolute fluctuat~on .. ~ Nt  Hi = (N1  Ni) may be restricted in magn1tudei for the contributions of terms corresponding "to large flucuations are in any case cut dovn by the d1stri1:ut1on factors ... /",...JI':/ W(N!>. ' The situation her~ is entirely similar to that we had ~en
"'~~~.....r:~ ...
, ~...... ~ ....... '.. ~~.....;. &_.... _... /~
'1
; .

85
discussing the energy fluctuations i~ the interaction of a closed system with a thermostat. . \:,J.::. aa.ve thus to compare each Zni Of!'> with its average value ~ (Ni) for relatively small diff~rences
i .
N!  NJ , and in order first to obtain the dependence of. the
1..... .
latter fUnotion on the average number WI we sh~ll make use of the
fact that log. Za (Ni) is directly oonnected with the free energy of the system Ri~' The following argument is essentially due to Tolman*'.
To present the matter as simply as possible, we shall
first consider in general terms the question of how the macrosco~. pic variables describing a system of identical elements depend. on
~the number of these elements. From this point of view the variables fall into two classes. Some of_them called intens1v~ ones, are proportional to the number of elements: they describe those properties of the homogeneous system which are, due to additive cantributions from th~ various parts 'of it; examples of ext~nsive variables are the voluee, the mean energy, and the characteristic functions: entropy. and free energy. Extensive variables give rise to corresponding "dens1tiesll, which are again intensive variables: the density 1,s the partial derivative of the extensiv~ variable with respect to the number of elements, provided that "the other independent variables kept cons tarrt in the' derivation, are Jll intensive.
It 1s always possible ~o have all inOcpendent variables intensive, f0C each pair a,A consists of an extensive and an intensive variable, ~ther of which_ can be chosen as repre~entative of the corresponding pr.ysical phenomenon. Thus, we may take either,t~e vol~ "or the pressure as the mechanical variab2e of a body of isotropic, structure; there are accord1nglytwo different
ufree enerkies" F(S, V) and G(6, p) which 'are usually associated in a rather arbitrary fashion, with the names of Helmholtz and Gibbs, respectively. In general, the passage from the characteristic fWlCi"ionF(9, a) to the £'unction G(S, A) is affected by a uLegendre transfo:rma t:ionll :
~~~
* cr R. Tolman, Phys. Rev. 51., 1160, 1940
................. :to.
86
G{9,~) = F[a, a{A)]  a{A) • b ,
where a{~) is· the inversion of the function A{a) computed from F(Q,a). Thus, G{U; e,p) = F{N; a,V) + Vp, and the corresponding
,
density is
g{9, p) = (~G) ~ P
lJ, N 'C1;
. (6)
so that
f .
. G{N; a, p) = N g(9,p)
,
(7)
From the connex1on
log Z (N; 9,p) =  P G (N; ·a, p)
between the statist+cal functiOO·log Z and the characteristic function G we derive immediately
log Z (N; S, p) = N f g{e, p)
. i.e._ log Z is an extensive function. This f'ormula, it 1s true, . is not strictly applicable to the case of our reservoirs, because the variation of the number of molecules in such a rese~oir does not occur in an e~terna~y controlled way, at constant pressure, but as the result of random flUctuations. The pressure in the reservoir 1sregulated by the mechanical equilibrium conditions across the scmiperm1allle membrane through ~lhich the exchange
of molecules with the system S takes place; its value is thus a function of the ave~age numbers NI, NZ of molecules in the system S, and every deviation from these .~verag~s brings ·about a perturbation ot·the equilibrium. For sm.all deviations, however, we may write:
=(N1  Ni) P gi{6,P1) ,
whe~e the reservoir pressure Pi 1s a function of the pressure R of the system S and of the~average number~ Nl, NZ •. Apart trom a factor indepen~ent of Ni, and ,.which therefore can.cels out. in the average (5) of the pPase function r, we have the required asymptotic· expression
;?"'.
Combining this with the factor W(Ni), given by (2); we may say that in the average (5), each number Ni is weighted by a factor of the
form 1 p ~ iNi
e
IN t 1
i
where the function ~ 1 (e',p; N1,NZ) = &1 + e log ifi represents the reservoir Ri in the same way as the parameter 9 represents the thermostat. In view of its fundamental importance for the determination of chemical equ1libria,. it is usually called the chemical
•
potential of the molecular species 1 in the system S.
The average (5) of the pha~e f'unct;1.on t now takes the
for~ 1 P(~Nf~~Z."'l2f· , ( ) PH(PNlNZ) ~~ Nllf>
~r" e l' P e· '/
: N N~iJN  1 NIN2 .
t= :'~ ~_ ~t~ ,I{l +'f .Jlz) . . , .'
~ ~:r e ~  Z (N1,NZ)
Nl:t2 • 1 2 .  . ,
'.
(8)
The phase integraticns indicated by f'WlN2 a~e over specific
phases, and the factor (Nll NZ1)1, has the effect of reducing them to integrations'over generic phases, in which each physical configuration of the system is counted only once. The introduction of this re_duct1on factor was presen,ted by Gibbs as a matter of convention, jUstified by its success, and it 'has given rise to much c~_~sed discussion: the present derivation*4 shows how nat~elly it comes in when 'a phYSical point of view ~s adopted,
in contrast to the more abstract attitude of Gibbs. According
to formula (8) the average_! can be described as a canon1~al average over generic phases for each possible set of numbers of molecules, weighted by an exponential distribution factor
c. [P(~ INI + ~ ZNZ) there is a complete anal!ogj' Wi~h~ t1io OonOep~
~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~
::. on' this point even Tolman IS paper which .we followed in all ' , ~"
othEr particulars is not sufficiently explicit.
t It is easily found (see: Tolman's paper) that the effect_of _
the perturbation of the equillbriUlil is or the order {lill N)/Ni • ",.~
' . ...
~ .
ot canonical average itself which is an ergodic average over each ~llcrgy surface, weighted by an exponential factor exp [ j3H]. ...., Hence the somewhat awkward nomenclature' proposed by Gibbs: the ordinary canonical distribution of copies of a closed system forms a
,
llpet1t ensemblell; the distr1but~on just found for an ~ system
is a IIgrand ensemblell• It 1s often, more simply called a grand canonical d1stribJtion.
3. Thermodynamics of open systems.
The statistical interpretation of the thermodynamical functions for open systems 1s a straight forward extension of the theory of closed systems, but it~adds an aspect of fundamental importance in chemistry, Viz, the way in which the characteristic functions depend on the nUmbers of molecules of various species present in the system.. ~le are here concerned, of ccur se , _ with
the equilibrium states of the system, _and therefore with the aver~ numbers of molecules. Let .us assume, for definiteness, that the independent meqhanical ~riable 1s the pressure; the characteristic function will thus be the Gibbs free energy.
Let us put
~ (Nl, NZ)
For a system with fixed numbers the tree energy by
(9)'
i
may define
G(Nl, NZ) =  e log,7; (N1, NZ):
this differs trom our previous definition only by a constant
(10)
,
ter. without interest so ~ong as the numbers ot moleculea do not
vary., Let us now consider the denominator of our fundamental formula (8):
(11)
and differentiate it With respect to all macroscopic variables: this differentiation includes the average numbers N1, HZ' but
leaves out the number.s N1, NZ' over which.8 summation 1s performed.
vie get , ...... ~~
;.1"
...
. 89
[d lOa(N1NZ)+Nl,d(P ~1)+ + N e d (P ;'2) )
where the aVe~ages refer to the distribut10n of the numbers N1,N2 given by the probability law'
n r 13(~ aN, + llN~)
Q .oZJ (Nl N.c) e  . ....
.'
The last equation may also be written
d (n + J3~1 Nl + fl~2 HZ) = d [fl . G (N1Ji2)J + ·.~ild ,N1+ P'~ed !fz{' c
" .
which shows that the characteristic fUnction in. terms of the inde~ pendent variables 9, P9. ifl.' NZ is .0+ fl (~l Nl + "(z.Nz.) •
It, in the right hand Side of equation (11) we replace the sum over the Ni's by the single term corresponding to the values Nl, NZ of these n~bers, we derive trom it an approximate form of the charact~r1stic function,
.o, J3 (~l NI + ~z HZ) ~ j3 G(N1. NZ) (13)
according to (10). To this approximation, which 1s justified provided that the' numbers N1, HZ are very large, we thus see that the characteristic function is still the free energy *, taken tor the average numbers of molecules, even when these average numbers are varied independently. Now, however, we see that it is quite essential to introduce 'the factor (N1INz.l )1, ''1hose role is. to reduce the II sum. over sta te s II to the physically distinct states or the ". ~ . " system, represented by "generic phases. Indeed, it i~ ,?~ly when
the integration is reduced to generic' phases that the f'ree energy
" ,
~
* The factor .~ can evidently be retained or omitted without Chany ing the property of t~e· function of' be"ing a "characteristic!! one •.
. .

~f 3 system of identical elements preserves its extensive character '1,0~;.~:'" ~, subdivision of the system into parts is taken into consideration. For if we make such a subdivision into two parts containing NA and NB elements, respectively, and consider the sums over state Z (NA + NB), Z (NA), i(NB) , integrated over all specific phases
of the total Systi~ and of the two parts separately, we have
. .
__ (NA+NB) 1
Z (NA + NB) ZeNA) Z(NB)
NA 1 NBJ
~.
i.e. precisely
The argument 1s immediately extended to the more general case we are considering of homogeneous systems ccntainlng several species of molecules. We must only generalise the concept o£.e.xtens1ve and intensive functions: these will denote homogeneous functions at the numbers of mol~cules, of degree 1 and 0, resp~ctively.
Using the approximation,(13) we get from (12) the
tundamental relation
"2 =f~a:.~] 'r ~ ~ (14)
• ""4., . 4, P 1 N iJ .
the symbol [Ni} indicating that all Nk IS except Ni are kept constaDt in the derivation. The extensive character ot G allows'us to . write
G = ~ N ~ = ~ Ni ~ : , (~5)
1 ioN 1
i
Comparing (~5) with (13), we see that, to the approximation con'
s iderc~, n~O. This means that the denomina tor of' f'ormula (B) practically redUces to unity, which considerably simplifies this f'ormula.
An interest1ng £cature of the free energy for a mixture ot.molecules is that it ~oes not reduce to the expression for a
single species if the dif£erent kinds of mo~ecules are identified •. Take e.g. the s1mple'case of a.mixture of' two species o~ molecules/ in the ideal gas state. 11e have
· ... '.
91
13 G =  NI log Zl  NZ log Zz + ~ log Ni 1 i
If we identity the molecules Zl = Z2 = Z, the two first ~erms reduce to the expected.·form  (Nl + NZ) log Z, but the last one is different from log (NI + NZ)l This "Gibbs paradox" s~ows how essential it is' for a consf.s tcnt, treatment of systems of various kinds of elements that these elements be dist1nguished by ~continuous criteria. Of course, the selection of those marks
which will be used to distinguish different spec1es 1s a matter orconvention, to be decided according to the circumstances of the concrete problem at hand. Thus; in ordinary chemical reactions, isotopes must be treated ~s identical elements, vhereas in questions of isotope separation they will naturally be distinguished into different species by taking into consideration the mass differences
.
. which are neglectgd in principle for the definition of chemical
species.
4. Characteristic functions invclving the temperature. In the preceding section, we have chosen as independent variables" besides the tempera'bure , the pr.essure and the numbers
of molecules: we have seen that the corresponding characteristic function is the free energy G(e, P, N) in the sense of Gibbs. Always keeping the temperature as the independent thermal parameter, we have1n principle 3 other possible combinations for the mechanical and chemical variables, according as we choose the 'volume instead Qf the pressure and the chemical potentials instead of the numbers of molecules. It is easy to ccnstruct the characteristic functicns for all these cases, by applying the suitable Legendre transformations. To begin with, we can eliminate the choice of 6, p, ~ as independent variables, for the characteristic function would then become G  2: ifi '£;1' i.e. by (15), identically zero. There accordingly re.dmin three possibilities, viz.*;:" ;
~~~~~~~~~~~~~~~~~~~~~~~~~
* The fUnction Jl in !16~ has a different meaning from the function denoted by.the same letter in sect10n 3. Henceforth, tl will always denote the function deCiDed in (16).
9!
.G (9, p, ii)' .= Lf. Ni i;i
F (8, v, N) = ~ Ni ~ i." P V. J.
.0.(9, V ~) =  p V
The property of being a characteristic function 1s expres
(~6)
sed in the three cases by similar equations:
dG = , ~d e + V dp + ~ g 1 d N i dF =  Sd e  pdV = + ~ ~ i d Ni d!l=  Sd e,  pdV  ~. Ni d'"'S i
Thus wee see that the chemical potentials can be defined by
"i'1 =( ~~i J e, v;\ ~ 11 (18)
just as well as by (14). 'On the other hand, we have
 (dO)
N1 =  a '$\1 e, v; { CS 1} (19)
'The statistical definitions of the characteristic functions n, F,G,' are ~ikew1se quit.c parallel: the present common feature ot celng simply related to the normalization factors of the respective statistical distribution·s. The. definit.lon of 1) is in tact
given by ~ 7. iN' _~ (P V) d '
~n e T 11 ( NlN2 •• ; I" Hl NZ··.
e =. N' Ul Ie. .. . . (20)
NlN2 ••• J: 1 2 .. • .. .1
(17)
w~lere the Hamiltonian is expressed in terms of .the volume as the' external mechanical variable; the proof is immediately cibtained* by computing the differential or~.with respect to the independent variables 9, V ·and the ~ i ' s , From (20) and (16) we derive for the function F the equation
eflF = L. e~~"s>i(N1iii)J ef: H (PNlNZ. •••. ;V) d
N11~2... NlI NZJ... (L N1NZ••• (Zl)
~~~~~~~~~~~~~~~~ 
* The situation is very similar to the above discussion of equation (11), but there we had chosen the pressure .as independent variable, which eventually led to the result that the other characteristic
function . .n.. defined by (11) vanishes. ./
r" .
• t ~ , _ • ,.,
~r we neglect the fluctuations of the numbers N~ around ~hc1r
?yc~~ges'Ni this ~xpre5sion reduces to
f3F
~ H{P:N N ••• ; V) 12
(22)
e .
Now, the ::orrasp0r,di , :; rigorous and approxinlatc expressions for
. the ft::.nct:'cn G aze tr .. c same as those' for F, except that the Hamiltonian must now be exprcsacd in tC:i.:'IJlS of p. This corrosponds , for the macroscopic qunntitics, to the passage from the lIencrgyll to the "enthalpy!! of the system. From the atomistic point, of viewt we
H (P, p) = H (P; V) + p V
,,~ ... , . '
have
the physical meaning of this relation is thQ~ ~n .this passage we change the def~itio~ of the mechanical system eonsideredl In fact if the volume is given, the Hamiltonian is simply the 'energy of tho system of molecules enclosed in a fiXod container.. If the pressure is given, we mus t i.r:::l.Ogine tha t, for ins tan co , one wall of the con» tainar is a ~ovable piston, upon which,the axternal pressure is exerted: this pitton is now part of the systam and contributes
a term pV to the Hamiltonian (where V is new regarded as a function of p and the other independent macroscopf.c variables.)