Professional Documents
Culture Documents
Abstract. The MaxEnt solutions are shown to display a variety of behaviors (beyond the traditional
and customary exponential one) if adequate dynamical information is inserted into the concomitant
entropic-variational principle. In particular, we show both theoretically and numerically that power laws
and power laws with exponential cut-offs emerge as equilibrium densities in proportional and other
dynamics.
2 Preliminary matters defined that follows a dynamical equation, like that of (2),
for example, the possible equilibrium situations towards
Our system consists of n abstract “elements” character- which the system may evolve can be obtained via MaxEnt
ized by a given “size”. Let us define by (i) expressing the entropy in terms of this dynamical
coordinate; and (ii) employing appropriate constraints.
(i) n as the total number of “elements” one is dealing The Shannon entropy measure in terms of x is defined
with; (up to a constant) as
(ii) xi as a positive-definite “size” assigned to the ith
element; S=− dxp(x) log p(x), (5)
(iii) x0 and xM as the minimum and maximum possible Ω
sizes;
(iv) p(x) as the relative number of elements with a size where Ω is a “volume” in x−space defined by the lower
given by x; and upper limits for the allowed sizes (x0 and xM , respec-
(v) x as the average-size. tively: x0 ≤ x ≤ xM ). Accordingly, Ω = xM − x0 .
Considering the continuous limit of the distribution p(x), The simplest scenario. We attempt first to use MaxEnt in
the conservation of both x and n guarantees the simplest possible instance, with just the normalization
xM xM constraint. The ensuing variational problem becomes
dxp(x) = 1 and dxp(x)x = x. (1)
x0 x0
δ S−μ dxp(x) = 0, (6)
Ω
Our goal here is that of finding out, via MaxEnt, the ex-
plicit form of p(x)dx from either some simple expectation- where μ is the associated Lagrange multiplier. The den-
values’ constraints or, and this is the novelty, from dynam- sity that extremizes this quantity is the constant one
ical information not of that kind. p(x)dx = Z −1 dx with Z = Ω the normalization constant
(and also by definition the partition function [4]). Remark
on the similarities of this scenario with that of a 1D clas-
sical ideal gas, with x representing the absolute value of a
3 Theoretical approach
position. Our example also describes an ensemble of non-
interacting particles of constant density represented by a
3.1 Brownian motion and statistical analogues gaussian distribution of velocities. We speak here of statis-
tical analogies between our original problem and physical
We first consider, as a control case, the dynamics behind ones.
the evolution of elements of sizes x via the linear equation:
Adding one additional constraint. In a second scenario we
ẋ = k (2) add a constraint on the first moment (the mean value
of the sizes) x = x, being thus led to the variational
where the dot denotes a time derivative, and k involves problem
the derivative of a Wiener process i.e., ki (t)kj (t ) =
Kδij δ(t − t ), with K the variance of k. We are in the
presence of the well-know Brownian motion, which obeys δ S−μ dxp(x) − λ dxp(x)x = 0, (7)
Ω Ω
the diffusion equation
where λ is the new associated Lagrange multiplier. The
∂t p(x, t) = D∂x2 p(x, t), (3) concomitant distribution is the well-know exponential
density
with D the diffusion coefficient and p(x, t) the non-equili- exp(−λx)dx
brium density at instant t. Starting at t = 0 with a Dirac- p(x)dx = . (8)
Z
delta distribution p(x, 0) = δ(x − x ), the solution to this
equation is a gaussian distribution of the form Note that xM could diverge here, but x0 has a lower
bound. The constants Z and Λ = x0 λ (defined for con-
1 (x−x )2 venience) can be obtained according to
p(x, t) = √ e− 4Dt . (4)
4πDt
e−Λ
Z = x0
Without constraints, the system evolves according to this Λ
distribution. At this point we introduce our main hypoth- (9)
esis, that will be amply validated below via numerical sim- −1 x
Λ +1= .
ulation of the equation of motion’s workings. x0
Central hypothesis: consider an ensemble of elements that We depict in Figure 1 an arbitrary case for x0 = 1,
evolve with a specified dynamics. If a coordinate can be n = 100 000 and x = 2.5 (Λ = 1/24). Let us assume
Eur. Phys. J. B (2012) 85: 147 Page 3 of 8
equation,
0.6 1
ẋ = kx (11)
0.5 0.1
where k is again indicative of the derivative of a Wiener
process. We now proceed to linearize the dynamic equa-
0.4 0.01
px
0.3 0.001
u̇ = k. (12)
0.2 104
0 5 10 15 20
In analogy with the precedent case, we have the usual
0.1 x
Brownian motion for this variable u, that obeys a diffusion
equation of type (3) with u instead of x. As for the x
0.0 observable one has
0 5 10 15 20
x
∂t p(x, t) = D∂x (x∂x (xp(x, t))) , (13)
Fig. 1. (Color online) Equilibrium density for the q = 0 case
(see Eq. (22) in the text) for x0 = 1, x = 2.5 and n = 100 000 solved (with an initial Dirac-delta p(x, 0) = δ(x − x )) via
(Λ = 1/24). Black line: prediction from the MaxEnt, blue dots:
the log-normal distribution
histogram of the asymptotic distribution of brownian walkers.
The shadowed area in the inset around √ the line reflects the 1 log2 (x/x )
numerical error of the simulation O(1/ n). p(x, t) = √ e− 4Dt . (14)
4πDtx
q-exponential [15–17]
0.4 104
1.0 1.5 2.0 3.0 5.0 7.0 10.0 15.020.0 30.0 x1−q −1
logq (x) = 1−q , if x > 0 (23)
x
0.2 1
expq (x) = [1 + (1 − q)x] 1−q , if 1 + (1 − q)x > 0, (24)
Fig. 2. (Color online) Same as Figure 1 for the q = 1 case of u = logq (x/x0 ). (25)
equation (22).
One finds in this fashion
We introduce now a second constraint, expressed in terms u̇ = k. (26)
of u via x = x0 eu = x, so that the Jaynes-MaxEnt
extremization problem becomes As before, this equation describes a Brownian motion and
thus a diffusion equation of type (3) in u. The equivalent
δ S−μ dup(u) − Λ u
dup(u)e = 0, (18) for x is of the form
Ω Ω
∂t p(x, t) = D∂x (xq ∂x (xq p(x, t))) , (27)
where we have used the definition Λ = x0 λ. We obtain
the density p(u)du = Z −1 exp(−Λeu )du. Changing back whose solution for an initial Dirac-delta p(x, 0) = δ(x−x )
to the observable x we arrive at is the q-log-normal distribution
1 exp (−Λx/x0 ) 1 (log q (x/x0 )−logq (x /x0 ))2
p(x)dx = dx. (19) p(x, t) = √ e− 4Dt . (28)
Z x 4πDtxq
The pertinent constants are obtained from the constraints
in the fashion We keep referring to our central hypothesis and using
Z = Γ (0, Λ) Shannon’s entropy in terms of u in the MaxEnt approach
e−Λ x (20) for finding equilibrium density:
= ,
ΛΓ (0, Λ) x0
S=− dup(u) log p(u). (29)
where Γ (a, z) is the so-called incomplete Gamma func- Ω
tion. This is then the expected equilibrium distribution for
a scale-free system, as those encountered in “opinion clus- The first constraint (normalization) is expressed, as usual,
ter dynamics” in networks with fixed number of nodes [11]. via
We display in Figure 2 the present scenario for the same δ S−μ dup(u) = 0, (30)
values of x0 , x and n as in the preceding section (now with Ω
Λ = 0.360743).
which yields a constant density for u as p(u)du = Z −1 du,
Assume again that our constrain is a mean-energy-one.
with Z = Ω = logq (xM ) − logq (x0 ). Now, changing to the
We can associate then to the distribution (19) the effective
proportional-growth Hamiltonian [4] observable x = x0 expq (u) we obtain
0.7 1.0
1.000 1.000
0.500 0.500
0.6
0.100 0.8 0.100
0.5 0.050 0.050
ΕΣ
ΕΣ
0.010 0.010
0.005
0.6 0.005
0.4
pΤ,u
pΤ,x
0.001 0.001
0.3
0.4 0 5 10 15 20
0 20 40 60 80 100
0.0 0.0
2 4 6 8 10 0 1 2 3 4
x u
1
Fig. 4. (Color online) Diffusion of the initial delta distribution
peaked at x = x = 2.5, at τ =0.2 (blue), 1.5 (green) and
4 (red) MC steps, with the asymptotic distribution (black)
compared with the MaxEnt prediction (black smooth line). 0.1
Inset: convergence of the standard deviation for the q = 0 case
(see
√ text). The shadowed area here reflects the numerical error
pΤ,x
1/ n. 0.01
1.2
(ii) we randomly select the ith walker and generate a new
position as xi = xi + kxqi dτ ;
1.0
(iii) we now correct the mean value as x = x + xq Δ,
where now Δ = (x − x)/xq , which guarantees
0.8
obeying the dynamical equation – note that we ex-
plicitly recover the previous cases when q = 0 and q =
pΤ,u
0.6
1. We accept the changes if, and only if min(x) ≥ x0 ;
(iv) we repeat (ii) and (iii) iteratively until convergence 0.4
in the distribution of x.
We again recommend the use of the linearized variable 0.2
u = logq (x/x0 ) via
0.0
(i) we fix the values of x0 , n, and x to generate the vector 0.0 0.5 1.0 1.5 2.0
u with ui = logq (x/x0 )), ∀i; u
(ii) we randomly select the ith walker and generate a new 1 1.000
0.500
position as ui = ui + kdτ ;
0.100
(iii) we now correct the mean value as u = u + Δ. Here 0.050
ΕΣ
Δ has no analytical expression, and is obtained from 0.1 0.010
0.005
the equation x0 expq (u + Δ) = x using an iterative
0.001
Newton algorithm, reaching convergence in few steps. pΤ,x
0 5 10 15 20
We accept the changes if, and only if min(u) ≥ 0 and 0.01
Τ MC steps
max(u) ≤ 1/(q − 1);
(iv) we repeat (ii) and (iii) iteratively until obtaining con-
vergence in the distribution of u. 0.001
5 Conclusions
0.5
0.1 0.010
a way of proceeding that entirely agrees with Jaynes’ phi- 0.005
losophy [1–4]. Indeed, we see that effective Hamiltonians
pΤ,x
0.001
0
for the process at hand are also a result of the MaxEnt 0.01 0 5 10 15 20 25 30
technique.
Τ MC steps
The JM-procedure can in this fashion still keep
Shannon’s measure while at the same yielding almost any 0.001
functional form for the ensuing variational PDF, power
laws in particular.
1 2 5 10 20 50
x
This work was partially supported by ANR DYNHELIUM
(ANR-08-BLAN-0146-01) Toulouse, and the Projects FQM- Fig. 6. (Color online) Top panels: same as Figure 4 for the
2445 and FQM-207 of the Junta de Andalucı́a. AP acknowl- q = 1.5 case at τ =0.1 (blue), 0.5 (green) and 2 (red) MC
edges support from the Senior Grant CEI Bio-Tic GENIL-SPR. steps. Bottom panels: q = 2 case.
Page 8 of 8 Eur. Phys. J. B (2012) 85: 147