Proc. Nati. Acad. Sci.

USA
Vol. 76, No. 5, pp. 2114-2117, May 1979
Physics

Irreversible processes at nonequilibrium steady states

(Liapunov function/"excess entropy"/covariance matrix/fluctuation-dissipation relation)
RONALD FORREST Fox
School of Physics, Georgia Institute of Technology, Atlanta, Georgia 30332
Communicated by Terrell L. Hill, February 23, 1979

ABSTRACT It is shown that a Liapunov criterion exists which the fluctuations in the thermodynamic variables arise.
for the stability of nonequilibrium steady states. This criterion On the average, each AM(t) is zero, and, if it is assumed to be
is based upon the fluctuation-dissipation relation, as was first Gaussian, then it is fully characterized by its correlation
pointed out by Keizer. At steady states, the Liapunov function strengths with other fj(t)s:
is constructed from the covariance matrix for the thermody-
namic variables. Unlike the situation around equilibrium, at ft ( _
steady states the covariance matrix and the "excess entropy" JO(fi(t)fj(s))ds = Qij [4]
matrix are not equivalent. The excess entropy, which serves as
the Liapunov function around equilibrium, does not work in this Qqj is symmetric and positive-definite. ( ...) denotes aver-
capacity at steady states. Keizer's Liapunov function must be aging.
viewed as the first correct candidate for a proper Liapunov The Boltzmann-Planck formula (8) with an entropy given
function for steady states. as in [2] yields the Einstein fluctuation formula (9) for the
The description of irreversible processes in the linear dynamical equilibrium fluctuations (10).
domain near full equilibrium was established in a series of pa-
Wi(d) = I(I 1
2

pers by Onsager (1, 2) and by Onsager and Machlup (3, 4). exp[- /2aiEijajI [5]
Stochastic time evolvement equations were used to describe the
fluctuations of the thermodynamic variables as well as the be- in which IIE 1 denotes the determinant of the entropy matrix.
havior of their averaged values. Both variational (1-3) and This stationary equilibrium distribution should be obtained
Liapunov (5) formulations of this near-equilibrium theory were asymptotically in time if the correlations of the thermodynamic
developed. variables described in [31 are computed. In fact, such a com-
What to do far from equilibrium was not determined by the putation yields
near-equilibrium theory. A similar mathematical framework
is now known to exist for the description of far-from-equilib- lim (ai(t)aj(t)) = (E- ) [6]
t -0 O

rium steady states (6). There are many similarities between the
thermodynamics of steady states and equilibria. A Liapunov provided that the fluctuation-dissipation relation
formulation is also possible for these nonequilibrium steady GE-' + E-'Gt = 2Q [7]
states (6). In both domains, the fluctuation-dissipation theorem
(7) plays an essential role. holds.
A Liapunov formulation is available for this stochastic process
ONSAGER'S THEORY (5, 6). The quantity, S(t) - So, serves as the Liapunov function.
The formulation of Onsager's theory to be presented here is Using [2], it is seen that
based upon a generalized representation, represented mathe- L(t) = S(t)-So <0 [8]
matically by the class of stationary, Markov, Gaussian processes for all values of d. From [1] and [2] it follows that
(7). The thermodynamic variables are denoted by as(t) for i =
1,2, . . . N. The ai(t)s satisfy linear rate laws d d
dt L(t)
= S(t) 1/2kB[ak(t) GkjEj1aj(t)
dtk =

d
dt as(t) =-E Gijaj(t), [1]
+ ai (t)EijGjkak(t)] > 0, [9]
in which Gij is the dissipative coupling matrix. The entropy, because the matrix
S(t), of the system is given by a quadratic form
GtE + EG = 2EQE [10]
S(t) = So - /2kBai(t)Eijaj (t ), [2] according to [7], and this form is manifestly positive-definite
in which E11 is a positive-definite, symmetric entropy matrix because Q is. Properties [8] and [9] of the Liapunov function
and kB is Boltzmann's constant. Actually, [1] is only the equation imply that the equilibrium state is asymptotically stable with
for the averaged thermodynamic variables, and the complete respect to small fluctuations (11). The heart of the Liapunov
set of equations is the stochastic process (7) property in this theory is the fluctuation-dissipation relation
d ai(t) = - E Gijaj(t) + fi(t), [7].
[3]
dt ~~~~i NONEQUILIBRIUM STEADY STATES
in which fi(t) is a random, or stochastic, driving force from A system may become transformed by applying imputs and
The publication costs of this article were defrayed in part by page outputs to it. It is possible to drive a system out of the equilib-
charge payment. This article must therefore be hereby marked "ad- rium state into a nonequilibrium steady state in this way. Such
vertisement" in accordance with 18 U. S. C. §1734 solely to indicate a steady state may be described by the system's thermodynamic
this fact. variables, the a(t)s. Away from equilibrium, the aj(t)s no longer
2114
 Physics: Fox Proc. Natl. Acad. Sci. USA 76 (1979) 2115
satisfy [3]. Analysis by the master equation technique of van Keizer (6) has proved that for an HsS and Rss with the prop-
Kampen (12) and Kubo et al. (13) and the limiting procedure erties stated above C-' is positive-definite. Therefore, the in-
of Kurtz (14, 15, 16) leads to a description (10) in which the equality
a1(t)s are decomposed by L(t) < 0 [231
ai(t) = -(t) + V-'/2Mp(t), 11] is valid. From [17] it follows that
in which as(t) is the "deterministic" part, whereas pi (t) is the d L(t) =-1/2[(Pk)H"(C-I)j;(Pj)
fluctuating deviation around the deterministic portion. The
as (t)s satisfy + (Hi)(C ')jjHik(gk)1] > [. 24]
dt di(t) = Ki[a-(t) . a-(t)j, [12] because the matrix -(HSStC-' + C-lHss) = C-lRssC-
according to [19], and this form is manifestly positive-definite
in which Ki is a nonlinear function of the ad(t)s. The ,ui(t)s because R S is. Again, it is the fluctuation-dissipation relation
satisfy [19], which is the core of the Liapunov property.
d N_ AN EXAMPLE
pi (t) = E Hijttj(t) + f1(t), [13]
Homogeneous phase chemical reactions provide the mathe-
in which Hij is defined by matical prototype for all considerations of specific examples.
The assumption of homogeneous phase means that only time
Hij--(aj K2 [14] dependence of molecular species populations is considered, and
and the fluctuating driving forces, fI(t), are correlated by all spatial variations are neglected. Nevertheless, the mathee-
matical apparatus employed in handling this relatively easy
rt problem is sophisticated enough to permit the treatment 'qf
J (f(t)fj(s))ds = 1/2Rj(iij,(t). .aN(t)]. [15] more difficult problems involving spatial dependence (17).
This correlation matrix, Ri2, unlike the Qqj in [4], is time de- Only modest changes in the mathematical structures are re-
pendent through its dependence on the ai(t)s. The necessity for quired to achieve such extensions in applicability.
a limiting procedure, such as the extention of the central limit Consider the elementary autocatalytic reaction involving
theorem provided by Kurtz (14, 15, 16) is manifested in the reaction species X and Y coupled to an input reservoir of
nonlinear as dependence of Ki in [12]. strength R y for species Y and to an output reservoir of strength
At a steady state, (d/dt)(diss) = 0. Therefore, Rx for species X. The mass action equations for this systemp
are
Ki[aIss ... aNSSI = 0 for all i. [16] k1
The fluctuating deviations, pi(t), now satisfy Ry -* Y (25]
k2
d hi()= E3 H t1yj(t) + fs(t), [17] X + Y 2::±2X
k3
in which HW = (6/d6a`s) Kil[dss ... aNj and kXR
X --R x.
f At
(fs( )fSS(s))ds = 1/2R [d8.a-.s..-s 1 [18] kX and k4 are independent of volume, whereas k2 and k3 are of
RUg is manifestly positive-definite, and, if the steady staLte is the form K2/V and K,3/V. Denoting the populations of species
X and Y by these same symbols, X and Y, yields the time-de-
asymptotically stable (6), then -H~s possesses positive real Iparts pendent equations
in each of its eigenvalues. The correlation matrix for thee19]s
satisfies X = k2XY - k:3X2 - k.X (261
dt
-H8sC - CHsst = RSS [191 d Y = kiRy - k2XY + k3X2.
in which dt
The input-output steps are strictly irreversible here because
Cij = tlim
emc
(Yi(t)Pj(t)). [20] it is the steady-state behavior of the autocatalytic reaction that
Eq. 19 is a fluctuation-dissipation for steady states (6, 10). is to be studied below. If, instead, reversible input-output steps
Kurtz' work justifies characterizing the fiS(t)s as Gaussian are used, steady-state behavior will not arise, but rather only
fluctuations. This property leads to a distribution function for a new equilibrium behavior for three coupled reactions will be
this process given in ref. 10 obtained.
A McQuarrie master equation (10, 18) may be constructed
(IIC 111)1/2 exp(-'/29i (C ')ijAj, [211 for this process. If the reaction takes place in a volume V, thenf
the variables that are directly suitable for application of Kurtz'
in which C-' is the inverse of the correlation matrix and par- limiting procedure (14, 15, 16) are given by
allels the E in [5]. However, at the steady state, C-l is no longer X
equal to E as it is in equilibrium, which is exhibited in [6]. = V = V R = Y and Rx = X.
Nevertheless, a Liapunov characterization of this process is still In these variables the McQuarrie master equation is
possible. The Liapunov function chosen below is to be con-
trasted with the candidate for this function proposed earlier a
P(x,y,t) = Vk1R P Xy-V't,
(5).
Choose as the Liapunov function the quantity
+ Vk4 ) Px +
(+ yj.
L(t) = -'/2 (i) (C- 1)ij ( ) [221
2116 Physics: Fox Proc. Natl. Acad. Sci. USA 76 (1979)

+ VK2 (X -) (I + (P X- V, Y + ) K(2) = K(2) =|-K2XY -

K3X X V

+ VKXX PX + Y V )
K(2) = kjRy + K2XY + K3X X-VJ)IV
-V(kRy + k4x + K2XY + K3X X-V P(x,y,t). [28]
The fact that x and y are independent variables here guarantees
HX)=-k44)
+ K2Y K3 (2x -

the definiteness properties of the matrices HI' and C which H(-)= K2X
appear below and which were discussed in the preceding sec-
tion. =H(l) K2Y + K3 (2x -V)
Let the vector fi denote the 2-tuple (x,y), and denote the
transition distribution by W(iii'), with the definition H =)-K2X. [34]
W(Oi,') VkRybx by y-,/v The steady-state values of x and y are determined according
+ Vk4x'bx'x+ 1/V5y'y to 1161 by
+ VK2 X'Y'6x'x-1/V6y'y+ 1/V K(')lxssyss] = 0 A = xy 135]
+ VK3X' X -
X'X+ 1/Vby'y-1/V, [29]
and these equations yield, in the thermodynamic limit V -,
Ry -X a, and RY is constant
in which the 6s are Kroecker deltas. For example,
=kk4 R
and ys k4 + Kl± R
y
K2 K2 k4
[36]
1 for x' = x--
1x'l/V=V Inserting these values into the expressions for H(,',? and K',?
O otherwise. [30] given in [34] provides the steady-state matrices needed in
equations 17 through 19. In particular, H", is generated by H('),
Using the transition distribution, Eq. 28 may be rewritten in and R", is identically given by limv- VK(2)ss.
the more compact form Eq. 19 determines the correlation matrix for the fluctuating
deviations from the steady-state values xss and ySs. Because the
6~P(n,t) = fd2n'[W(fi,n')P(fi',t) W(niji)P(ii,t)]. [31 ]
- correlation matrix Cvns is symmetric, there are only three in-
dependent matrix elements, and the solution to [ 19] is obtained
W(ri,ii') is not symmetric in fi and i' as is seen by substitu- from the equivalent 3 X 3 matrix equations:
tion
W(ii',ri) = VkRybxx byy-l/v ( 2HXX 0 ss
R (2)ss
2HSs 2HSS
+ Vk4xSxx'+ I/Vbyy' 0 C =-R(2)ss
yy
+ VK2XYbxx-I/Vbyy'+ 1/V
rss
Hlyx HSxs 1SxX +
HSJ I RCJSx3y7
+ VK3X (X - )I bxx+ 1/Vbyy'-1/V. [32] [371
By using 1341, an explicit solution is obtained from:

\Cxx
= -[4(det H)(trace H)]-l 2HssxxH
(Hssxx +
2(Hss )2
Hss)/}
xy
- 2Hss Hss
+4HssHss 4HSS 38J
[381
SsXY /R (2)ss
-~~4H Hss \(R,22S
R-)s
Cxy -2HssYXyHss
x y -2Hss Hss 4Hss Hyss xy

The application of Kurtz' method (10, 14) requires computation At this stage, the question to ask is: Is the solution for CM,
of several quantities that depend on W(',fi): obtained from 1381 identical with the "excess entropy" evalu-
ated relative to the steady state as has been proposed by Pri-
K(1)= d2n(n - n,)W(n',n ) ji = x,y gogine (5)?
Kp,? -- fdn'(nW - n,)(nf, - n,,)W(i',fi) The entropy expression given in 121 is the quadratic ap-
a
proximation to the actual entropy which is given in ref. 19.
)W
H (1,)=-
H K(1)
' On,Ic
[33] S(t) = So - kB(x In Vx + y In Vy). 139]
The expli t values for these quantities in the present situation Around the steady state, S may be expanded
are S = Sss + hS + 1/2(2S + ...
K -1)= -k4x + K2XY- K(X V)X = Sss - kB(x In Vxss + biy In Vyss)
- 1/2kB [(X)2+ (y2 +---- [40]
K(-)= kljR-K2XY + KS x-44I x
Eq. 40 may be expressed in terms of a steady-state matrix Es]
which is given by
KX= k4x + K2XY + K3X A
V44 V
1/2 52S = -'/2kB(bxE" hx + hyE" (5y), 1411
 Physics: Fox Proc. Natl. Acad. Sci. USA 76 (1979) 2117
where fluctuating forces is determined by the master equation limit
procedure and the fluctuation-dissipation relation determines
the covariance of the fluctuations of the thermodynamic vari-
( 0Evx) ( 1) [421 ables. This second point of view has been the basis for Keizer's
approach (6) to these problems, and it directly reflects the re-
0 ESS ~ yS
lationship between the underlying microscopic dynamics and
Yss the macroscopic thermodynamic description.
No off-diagonal elements of this matrix are non-zero. If this The proposal by Prigogine (5) that the "excess entropy"
result is to be in agreement with the solution exhibited by [381, provides a Liapunov criterion both near equilibrium and near
then steady states has appeared in several contexts (21, 22). The
|Cxx / Ess )-1 lass possibility that it has biological significance has been suggested
rather forcefully by Eigen (23, 24). However, Keizer's work (6,
CYY = (ESS)-I ys [431 20) suggests, as has been demonstrated in this paper, that the
"excess entropy" does not provide a Liapunov criterion near
steady states, and that, instead, the covariance of the fluctuating
Identity 43 is equivalent to the requirement that follows from thermodynamic variables does provide such a criterion. It has
1371 that been the purpose of this paper to bring these facts to the at-
|2HSS
(2Hz 0 lxssl
2Hy() IR ()SSI[4
I -[44]
tention of a much wider audience.
| 2HSy /ISS IR I (2 1.
2.
Onsager, L. (1931) Phys. Rev. 37, 405-426.
Onsager, L. (1931) Phys. Rev. 38, 2265-2279.
or equivalently 3. Onsager, L. & Machiup, S. (1953) Phys. Rev. 91, 1505-1512.
R (2)ss R (2)ss 4. Machlup, S. & Onsager, L. (1953) Phys. Rev. 91, 1512-1515.
XSS and yss =-.JU. .
xx [45] 5. Prigogine, I. (1973) in The Physicist's Conception of Nature, ed.
2HSS 2HSS Mehra, J. (Reidel, Boston), pp. 585-588.
Now, the expression for yss is indeed correct, but that for xSS 6. Keizer, J. (1976) J. Chem. Phys. 65, 4431-4444.
is not. Therefore, identity 43 is not verified. The "excess en- 7. Fox, R. F. & Uhlenbeck, G. E. (1970) Phys. Fluids 13, 1893-
tropy" around the steady state does not generate the covariance 1902.
of the fluctuations at the steady state. 8. Kubo, R. (1965) Statistical Mechanics (North-Holland, Am-
sterdam), pp. 51-52.
9. Einstein, A. (1910) Ann. Phys. 33, 1275-1294.
DISCUSSION OF RESULTS 10. Fox, R. F. (1978) Phys. Rep. 48, 181-283.
It is easy to verify that if the input and output terms in [25] are 11. Hahn, W. (1967) The Stability of Motion (Springer, Berlin).
omitted, then the "excess entropy" and the covariance of the 12. van Kampen, N. G. (1961) Can. J. Phys. 39,551-559.
fluctuations, both taken relative to the equilibrium state in this 13. Kubo, R., Matsuo, K. & Kitahara, K. (1973) J. Stat. Phys. 9,
51-96.
case, are in agreement. This is merely another way of expressing 14. Kurtz, T. (1972) J. Chem. Phys. 57, 2976-2978.
the results described in equations 1-10. Away from equilibrium, 15. Kurtz, T. (1970) J. Appl. Probab. 7, 49-58.
when input and output terms generate steady states, these two 16. Kurtz, T. (1971) J. Appl. Probab. 8, 344-356.
concepts are no longer equivalent. It has been shown above that 17. Keizer, J. (1976) J. Chem. Phys. 64, 1679-1687.
the Liapunov property for steady states depends directly upon 18. McQuarrie, D. (1967) J. Appl. Probab. 4, 413-478.
the covariance of the fluctuations and not upon the "excess 19. Landau, L. & Lifshitz, E. M. (1958) Statistical Physics (Perga-
entropy" (20). mon, London), 288-291.
The significance of the fluctuation-dissipation relation 20. Keizer, J. & Fox, R. F. (1974) Proc. Natl. Acad. Sci. USA 71,
around equilibrium and around steady states is quite different 192-196.
even though in both cases it is the key to the Liapunov property. 21. Glansdorff, P. & Prigogine, I. (1971) Structure, Stability, and
Near equilibrium, the covariance of the fluctuations of the Fluctuations (Wiley/Interscience, London).
22. Nicolis, G. & Prigogine, I. (1977) Self-Organization in Non-
thermodynamic variables is known from the Boltzmann- equilibrium Systems (Wiley/Interscience, New York).
Planck-Einstein formula, and the fluctuation-dissipation 23. Eigen, M. (1971) Naturwissenschaften 58, 465-523.
relation determines the correlation strengths of fluctuating 24. Eigen, M. (1973) in The Physicist's Conception of Nature, ed.
forces. Near a steady state, the correlation strengths of the Mehra, J. (Reidel, Boston), pp. 609-610.