See discussions, stats, and author profiles for this publication at: [Link]

net/publication/226805234

Possible generalization of Boltzmann-Gibbs statistics

Article in Journal of Statistical Physics · July 1988

DOI: 10.1007/BF01016429

CITATIONS READS

9,614 6,298

1 author:

Constantino Tsallis
Centro Brasileiro de Pesquisas Físicas
243 PUBLICATIONS 19,113 CITATIONS

SEE PROFILE

All content following this page was uploaded by Constantino Tsallis on 12 January 2015.

The user has requested enhancement of the downloaded file.

Journal of Statistical Physics, Vol. 52, Nos. 1/2, 1988

Possible Generalization of
Boltzmann-Gibbs Statistics
C o n s t a n t i n o Tsallis 1

Received November 12, 1987, revision received March 8, 1988

With the use of a quantity normally scaled in multifractals, a generalized form

is postulated for entropy, namely Sq==-k[1-Zwt pq]/(q-1), where q e ~
characterizes the generalization and {Pi} are the probabilities associated with W
(microscopic) configurations (WE N). The main properties associated with this
entropy are established, particularly those corresponding to the microcanonical
and canonical ensembles. The Boltzmann-Gibbs statistics is recovered as the
q ---, 1 limit.

KEY W O R D S : Generalized statistics; entropy; multifractals; statistical

ensembles.

Multifractal concepts and structures are quickly acquiring importance in

many active areas of research (e.g., nonlinear dynamical systems, growth
models, commensurate/incommensurate structures). This is due to their
utility as well as to their elegance. Within this framework, the quantity that
is normally scaled is pq, where Pi is the probability associated with an event
and q is any real number. (1) I shall use this quantity to generalize the
standard expression of the entropy S in information theory, namely
S = - k Z iw= l p i l n p i , where WEN is the total number of possible
(microscopic) configurations and {pi} is the associated probabilities. I
postulate for the entropy

Sq~k 1 - E w ' p]
q-1 (qE~) (t)

1 Centro Brasileiro de Pesquisas Fisicas-CBPF/CNPq, Rua Xavier Sigaud 150, 22290 Rio de
Janeiro, R J, Brazil.

479
0022-4715/88/0700-0479506.00/0 9 1988 PlenumPubLishingCorporation
822/52/1-2-31
480 Tsallis

where k is a conventional positive constant and Z W l p i = 1. It is

immediately verified that

$1 - lira Sq=k lira 1 - Z w l &exp[(q- 1)ln & ]

q~l q~l q-1
W
= - k • p~lnp~ (1')
i=1

where I have used the replica-trick type of expansion. Figure 1 illustrates

definition (1). One can rewrite Sq as follows:
k w
Su_ ~. p i ( l _ p q 1) (2)
q-li=l
which m a k e s evident that Sq ~ 0 in all cases. It vanishes for W = 1, Vq, as
well as for W > 1, q > 0, and only one event with probability one (all the
others having vanishing probabilities).
Microcanonical Ensemble. We want to extremize S u with the
condition Z W ~ p ~ = 1. By introducing a L a g r a n g e parameter, it is
straightforward to obtain that Sq is extremized, for all values of q, in the
case of equiprobability, i.e., Pi = 1/W, Vi, and consequently

H/q-q- 1
Sq=k 1 -q (3)

Sq
k

15

I 0

0.5 '~ 0.5

o 0.5 p"
Fig. 1. Plot of Sq({Pi} ) for W=2 and typical values of q (numbers on curves). Notice the
monotonic influence of q, a fact that reappears in a variety of properties.
Generalization of Boltzmann-Gibbs Statistics 481

It is immediately verified that

S 1 ~- k In W (Y)

thus recovering the celebrated Boltzmann expression. Figure 2 illustrates

Eq.(3). The Sq given by Eq. (3) diverges if q~<l and saturates [at
Sq = k / ( q - 1 ) ] if q > 1, in the W--* oo limit. It is straightforward to prove
that the extremum indicated in Eq. (3) is a maximum (minimum) for q > 0
( q < 0 ) ; for q = 0 , S q ( { p i } ) = k ( W - 1) for all {Pi}. Finally, Eq. (3) implies
Sq _ e (1 q) S l / k - - 1
(4)
k 1-q

Concavity. Let us extend here a property already mentioned,

namely that q > 0 (q < 0) implies that the extremum of S u is a maximum
(minimum). Let {Pi} and {p~} be two sets of probabilities corresponding
to a unique set of W possibilities, and 2 such that 0 < 2 < 1. Define an inter-
mediate probability law as follows:

pT==_2pi+(1-2)p; (Vi) (5)

Sq
m
k

2_ 0.5

0
i 2 3 4 W
Fig. 2. Value of the entropy at its extremum for typical values of q (numbers on curves). The
dashed line indicates the W--, oo asymptote of S2/k.
482 Tsallis

and also

Aq=-gq({p;'}) - [)~Sq({Pi})+ (l -)~) Sq({p;})] (6)

It is straightforward to prove that Aq >10 if q > 0, Aq~ 0 if q < 0, and Aq : 0

if q = 0. The equalities hold for q r 0 for p~ = p;, Vi.
AdditiviW. Let us assume two independent systems A and B with
ensembles of configurational possibilities g2A= { 1, 2 ..... i,..., WA } and g2e -
{1, 2,..., j,..., We}, respectively, the corresponding probabilities being {pA}
and {p~}. Now consider A w B, the ensemble of possibilities being g2A~ e =
{(1, 1), (1, 2),..., (i,j),..., (WA, We)}; let p ~ e denote the corresponding
probabilities. The independence of the systems means that p~ ~ s = p~p~,
V(i, j), hence

E (p~A ~ B ) q = (pJ)~ (pT) ~

i,j i 1

Hence [-using Eq. (1)]

~qA~e __ SA + SqS (additivity) (7)

with

Sq = k ln[1 + (1 - q)Sq/k] (8)

1-q

In the q ~ 1 limit, Eq. (7) becomes S~ ~ e = S A + Sf, thus recovering the

standard additivity of the entropies of independent systems. For arbitrary
q, Sq reproduces the Renyi entropy. (2)
To study the case of correlated systems [i.e., ,-0 n A~B is not equal to
(SZwA1 p A ~ e ) ( Z w , 1 p ~ e ) for all (i, j)], it is useful to define

rq({pJ~e}) sq
--A ~ B ({p~
A w B })-s~
--A pj~e Sq
-- -e PA~e
u

\kj=l i 1

It is clear from Eq. (7) that independence (no correlation) implies Fq = 0,

Vq. For arbitrary and fixed {p~ ~ B} implying correlation, it is easy to prove
that F 1 < 0 (subadditivity of the standard entropies of correlated systems)
and F0 = 0. For arbitrary values of q, Fq presents a great sensitivity to
{pj~B}, it might be positive or negative for q>> 1 as well as for q ~ - 1 ,
and typically exhibits more than one extremum. Extensive and systematic
computer verification indicates that, generally speaking, Fq varies smoothly
with q, but presents no particular regularities besides F 0 = 0 and F~ ~<0.
Generalization of Boltzmann-Gibbs Statistics 483

When {pA~,} gradually approach vanishing correlation, Fq gradually

flattens until eventually achieving Fq = O, Vq.
Canonical Ensemble. We want to extremize Sq with the con-
ditions Z ~ I Pi = 1 and
w
2 Pi~i : Uq (9)
i=l

where {e~} and Uq are known real numbers (the same value ei might be
associated with more than one possible configuration); they will be referred
to as generalized spectrum and generalized internal energy. I introduce the
and fl Lagrange parameters and define the quantity
Sq w w
Oq---~+~ ~ pi-c~fl(q-1) ~ Pi~i (10)
k=l i=1

which is written this way for future convenience. Imposing O~q/~Pi = O, Vi,
one obtains picc [ 1 - f l ( q - 1 ) e i ] I/(q 1); hence,
[ 1 - fl(q - 1)s~] 1/(q- 1)
Pi- (11)
Zq
with
W
Zq ~- Z E1 - - f l ( q - 1 ) ~ 1 ] 1/(q 1) (12)
l=1

It is immediately verified that, in the q ~ 1 limit, one recovers

pi=e fl~i/Z1 (11')

with
l,v
Z1- ~ e -~ (12')
l=1

It is straightforward to see that an alternative manner for obtaining

the power-law distribution expressed in Eq. (11) is to extremize Sq (or
equivalently Sq) with the condition Y~w 1 pqei= Uq [instead of Eq. (9)].
If A and B are two independent systems with probabilities (spectrum)
{pA }({e{ } ) and {pT} ( {e~ }), respectively, the probabilities corresponding
to A w B satisfy p~ Us = pAp~, V(i, j). This implies

1 -- fl(q -- 1 ) eijA u g = [1-fl(q- 1)eA][1--fl(q-- 1)ST] (13)

484 Tsallis

or equivalently

eo-A,~B=gA+gff (14)

with

ln[1 + [3(1 - q)e]

g- (15)
[3(1 - q )

In the q--+ 1 limit (and/or [ 3 ~ 0 limit), Eq. (14) becomes %.A~B--e(

+e~,-
thus recovering the standard energy additivity. The property (14), together
with the factorization of probabilities, placed in Eq. (9) yields

U u A ~ = C A + U~ (16)

with

ln[1 + [3(1 - q) Uq] (17)

Uq- [3(1 - q )

In the q ~ 1 limit (and/or [3 ~ 0 limit), Eq. (16) becomes U{ ~8 = U{ + Uf,

thus recovering the standard additivity of the internal energies of indepen-
dent systems.
I now discuss the main characteristics of the distribution law (11).
First, notice that this distribution is invariant under the transformation

[1 -[3(q- 1)et-] ~ [1 -[3(q- 1)et-]]-I -[3(q- 1)eo]

for all l, e0 being an arbitrary fixed real number. In other words, the
distribution (11) is invariant under gr [this is in fact a trivial
consequence of the fact that the distribution can be formally rewritten as
Pi ~: exp(-[3gi)]. For f l ( q - 1 ) - ~ 0, we recover the well-known invariance
of the Boltzmann-Gibbs statistics under uniform translation of the energy
spectrum. Figure 3 illustrates distribution (11). Notice that, for q > 1, pi = 0
for all levels such that s~>~l/[fl(q-1)] (~<~-l/[[fl](q-I)]) if fl>O
(fl<O), i.e., positive (negative) "temperatures." On the other hand, for
q < l , the levels such that e ; ~ < - l [ f l ( 1 - q ) ] (~>~1/[-[[31 ( l - q ) ] ) are, if
fl > 0 (fl < 0), highly occupied, in a way that is clearly reminiscent of the
Bose-Einstein condensation.
To better realize the unusual properties of the present statistics, it is
instructive to analyze the following situation. Assume q > l, fl > O, and {ei}
such that 0 < el < e2 < "'" < e w (W might even diverge). When 1/[3is above
( q - 1)ew, all levels have a finite occupancy probability; when
( q - 1 ) e w l<l/[3<(q-1)ew, then pl>p2> "" >pw_~>pw=O. The
Generalization of Boltzmann-Gibbs Statistics 485

Zq Pi

<.
q-- 9

-' o t z pEi
Fig. 3. The distribution law of Eq. (11 ) as a function of/~i. The curves are parametrized by
q: q = 1, standard exponential law; q > 1, the distribution pressents a cutoff at ]~ei= 1/(q- 1)
(with a slope of 0, -1, and - ~ for q<2, q=2, and q>2, respectively) and diverges for
/~et~ - ~ ; q < 1, the distribution diverges at /3e~= - 1 / ( 1 - q) (the dashed line indicates the
asymptote for q ~ 0) and vanishes for/3e~ +co.

probabilities successively vanish while 1//3 decreases. One eventually arrives

at ( q - 1)g~ < 1//~ < ( q - 1)g2, which implies p~ -- 1. Finally, the tem-
peratures 1//3 in the interval [0, ( q - 1 ) e l ] are physically unaccessible, thus
generalizing the nonaccessibility of 1//3--0 in standard thermodynamics. A
simple example will illustrate this and similar facts.

Application. Consider two nondegenerate levels with values el -=

-6 and ~ 2 = e + 6 (3 > 0 ; g-~0). The quantity Uq(/3) is given by Uq=
g~ Pl + e2P2. A straightforward calculation yields

[1 - ( q - 1)(g/6- 1)/x] l/(q- ~)- [1 - ( q - 1)(e/6 + 1)/x] 1/(q ~)

Yq = - [1 - ( q - 1)(g/6 - 1)/x] 1/(q- 1)+ [1 - ( q - 1)(g/6 + 1)/x] 1/(q- 1) (18)

with x = I/f16 and yq=(Uq--8)/c~E[ - 1 , 1]. E q u a t i o n (t8) is invariant

under (x, yq, q - - 1, g/c~) ~ (X, yq, - - ( q - - l), --e/b) and also under
(x, yq, q, g/c~) -.,,. ( - x , - y q , q, - e / 0 ) . Consequently, it suffices to discuss
q >~ 1 and g/6 >~O. In the limit q--, 1, one obtains Yl = - t h ( 1 / x ) , Vg/6. F o r
q ~ 1, yq(X) depends on e/6: see Figs. 4 and 5.
486 Tsallis

(o} (b)

0 (~:o) __o .... (o<~<11

I I r i i E
-3 -2 -1Jo| ,: 2 ~l.-x-
-,~
(c) (d)
2~ E -,- ~'

0
C~- ') --
E,-(S
~ .... (~->,.)

Y~T~- CloT- 1

-3 -2 - 1 0 ~ I 2, 3 X

.,p~/
Fig. 4. The q = 2 reduced internal "energy" as a function of the reduced "temperature" (see
text) for a nondegenerate two-level system and typical values of ~/6. The dashed region in (d)
indicates the unaccessible "temperatures."

I conclude by recalling that, using the quantity normally scaled for

multifractals, I have postulated an expression for the entropy that
generalizes the usual one (recovered for the parameter q ~ 1). By preserv-
ing the standard variational principle, I have established the
microcanonical and canonical distributions, as well as several other proper-
ties. Some of the emerging peculiar characteristics are illustrated through a
simple example. One of the most interesting is the fact that the unaccessible
"temperatures" might belong to a finite interval that shrinks on the
T = 0 point in the q --+ 1 limit. Finally, the fact that Sq/k, [3ei, and BUq are
additive under one and the same functional form {namely f ( x ) = _
In I-1 + ( 1 - q ) x ] / ( q - 1)} opens the door to the generalization of standard
thermodynamics through the introduction of appropriate generalized ther-
modynamic potentials. Applications of these generalized equilibrium
 Generalization of Boltzmann-Gibbs Statistics 487

Yc

(~/i~-- 0 )

l i i I L I
-3 -2 -I 0 I 2 3 X"

-1

Fig. 5. Reduced internal "energy" as a function of the reduced "temperature" (see text) for a
nondegenerate two-level system and typical values of q (numbers on curves).

statistics in physics (e.g., fractals, multifractals), information theory, or any

other branch of knowledge using probabilistic concepts would be extremely
welcome.

ACKNOWLEDGMENTS
I am very indebted to E. M. F. Curado, H. J. Herrmann, R. Maynard,
and A. Coniglio for very stimulating discussions. Computational assistance
by S. Cannas as well as useful remarks by S. R. A. Salinas, F. C. S~ Barreto,
S. Coutinho, and J. S. Helman are also gratefully acknowledged.

REFERENCES
1. H, G. E. Hentschel and I. Procaccia, Physica D 8:435 (1983); T. C. Halsley, M. H. Jensen,
L. P. Kadanoff, I. Procaccia, and B. I. Shraiman, Phys. Rev. A 33:1141 (1986); G. Paladin
and A. Vulpiani, Phys. Rep. 156:147 (1987).
2. A. R6nyi, Probability Theory (North-Holland, 1970).

Communicated by J. L. Lebowitz

View publication stats