Physics Letters A 326 (2004) 20–26

www.elsevier.com/locate/pla

Escort Husimi distributions, Fisher information and nonextensivity

F. Pennini, A. Plastino ∗
Instituto de Física (IFLP), Universidad Nacional de La Plata and Argentina’s National Research Council (CONICET),
C.C. 727, 1900 La Plata, Argentina
Received 3 February 2004; received in revised form 6 April 2004; accepted 8 April 2004
Available online 28 April 2004
Communicated by C.R. Doering

Abstract
We evaluate generalized information measures constructed with Husimi distributions and connect them with the Wehrl
entropy, on the one hand, and with thermal uncertainty relations, on the other one. The concept of escort distribution plays
a central role in such a study. A new interpretation concerning the meaning of the nonextensivity index q is thereby provided.
A physical lower bound for q is also established, together with a “state equation” for q that transforms the escort-Cramer–Rao
bound into a thermal uncertainty relation.
 2004 Elsevier B.V. All rights reserved.

PACS: 02.50.-r; 03.65.-w; 89.70.+c

Keywords: Fisher information; Husimi distributions; Escort distributions; Nonextensivity

1. Introduction supplying a nontrivial measure of uncertainty due to

both thermal and quantum fluctuations [4]. It has been
We will be concerned here with generalizations of shown in [5] that intriguing connections link W to
two important information-theoretic uncertainty mea- Fisher’s information measure I . We will here also
sures: those of Fisher’s (I ) [1,2] and Wehrl’s (W ) [3]. generalize these connections.
The Wehrl entropy verifies the relation W
1 [4], and The above referred to generalizations (indeed, non-
this bound represents a strengthened version of the un- extensive extensions [6,7]) of W and I , together with
certainty principle. A similar case can be made for I their associated uncertainty applications will be shown
[1,2]. to shed some light onto the meaning of the nonex-
In the case of a harmonic oscillator in a thermal tensivity parameter q. Establishing adequate q-criteria
state, W coincides with the logarithmic information still constitutes an open problem for nonextensive ther-
measure of Shannon’s in the high temperature regime. mostatistics, although great progress has been made in
However, it does not vanish at zero temperature, thus deriving from first principles the appropriate q-value
for special dynamical problems, some of them related
* Corresponding author. to Hamiltonian systems [6,8]. These systems are clas-
E-mail addresses: pennini@fisica.unlp.edu.ar (F. Pennini), sical ones, though. Our efforts here will be directed,
plastino@fisica.unlp.edu.ar (A. Plastino). instead, towards quantum systems.
0375-9601/$ – see front matter  2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.physleta.2004.04.024
 F. Pennini, A. Plastino / Physics Letters A 326 (2004) 20–26 21

dp dx
The Letter is organized as follows. In order to d2 z = d(z) d(z) = ≡ dp dx  . (4)
facilitate the reader’s task, some preliminary material 2h̄
is presented in Section 2. We start our present quest in Consider now a system characterized by a Hamil-
Section 3 by first generalizing the concept of Wehrl tonian Ĥ . Husimi has introduced an important distrib-
entropy to an nonextensive environment, obtaining ution function [4,10]
a “q-Wehrl” entropy that provides us with a new
interpretation for the index q and then studying a µ(p, x) = z|ρ̂|z, (5)
Fisher’s information measure constructed with what
we call “escort Husimi” distributions, which allows us associated to the density matrix ρ̂ of the system. The
to obtain a physical lower bound to the nonextensivity function µ(p, x) is normalized in the fashion
parameter q. Finally, some conclusions are drawn in  
Section 4. dp dx dp dx 
µ(p, x) = µ(p , x  ) = 1, (6)
2π h̄ π

2. Background material which makes it evident that x  and p are the “natural”
µ-variables, of (HO-ground state) variance unity. The
2.1. Coherent states and Wehrl information distribution µ is indeed a Wigner distribution smeared
over a phase-space region of size h̄ [4]. It is important
In [4] the authors discuss quantum-mechanical for our present purposes to remark that the Husimi
phase-space distributions expressed in terms of the distribution (HD) is a positive definite one [4]. The HD
celebrated coherent states |z of the harmonic oscil- may be thought of as a “classical” distribution over
lator (HO), whose Hamiltonian operator Ĥ is given phase-space [11].
by One of the main tenets of Information Theory is

 that one can associate an information measure to any
1 probability distribution [13]. The Shannon informa-
Ĥo = h̄ω â â + .
†
(1)
2 tion measure associated to the Husimi distribution is
The coherent states are eigenstates of the destruction called the Wehrl entropy W [3]
operator â, i.e., 
dp dx
W =− µ(p, x) ln µ(p, x). (7)
â = i(2h̄ωm)−1/2 p̂ + (mω/2h̄)1/2x̂, (2) 2π h̄
 
1 x ip
z= + As shown by Lieb [12], this special entropic form
2 σx σp verifies the inequality
 1/2
mω
= x + i(2h̄ωm)−1/2 p ≡ x  + ip , (3)
2h̄ W
1. (8)
with
x p 2.1.1. Canonical Husimi distribution
x = , p = , Let Ô be an operator relevant for the system’s
2σx 2σp
    description. The “thermal” mean value of Ô in Gibbs’
h̄ 1/2 h̄mω 1/2 canonical ensemble is given by [11]
σx = , σp = ,
2mω 2  
σx σp = h̄/2. Ô = Tr ρ̂ Ô ,

Variances σ are evaluated for the HO ground state. Co- ρ̂ = Z −1 e−β Ĥ , Z = Tr e−β Ĥ , (9)
herent states span Hilbert’s space, constitute an over-
complete basis and obey the completeness rule [9] with ρ̂ the system’s canonical density matrix, Z the
 2  pertinent partition function, and β = 1/kT , being T
d z dp dx
|zz| = |p, xp, x| = 1, the temperature, with k the Boltzmann constant, to be
π 2π h̄ set equal to unity hereafter.
22 F. Pennini, A. Plastino / Physics Letters A 326 (2004) 20–26

2.1.2. HO Husimi distribution observer makes a measurement of x and has to best in-
In the important HO instance (Ĥ ≡ Ĥo ), if we fer θ from this measurement, calling the resulting es-
denote with |n the HO-eigenstates, associated to the timate θ̃ = θ̃(x). One wonders how well θ can be de-
eigenvalues En = h̄ω(n + 1/2), one has [4] termined. Estimation theory [14] asserts that the best
possible estimator θ̃(x), after a very large number of
1
−βH  2
z|ρ|z = e z|n , x-samples is examined, suffers a mean-square error e2
Z n from θ that obeys a relationship involving Fisher’s I ,
  namely, I e2 = 1, where the Fisher information mea-
z|n2 = |z| e−|z|2 ,
2n
(10) sure I is of the form
n!   
∂ ln fθ (x) 2
entailing that I (θ ) = dx fθ (x) . (15)
 ∂θ
−β h̄ω )|z|2
µ(p, x) = 1 − e−β h̄ω e−(1−e , (11) This “best” estimator is called the efficient esti-
mator. Any other estimator must have a larger mean-
or, in terms of the “natural” variables p and x 
square error. The only proviso to the above result is
 −β h̄ω )[p  2 +x  2 ]
that all estimators be unbiased, i.e., satisfy θ̃ (x) = θ .
µ(p , x  ) = 1 − e−β h̄ω e−(1−e , (12)
Thus, Fisher’s information measure has a lower bound,
gives the HO Husimi distribution, which, after inte- in the sense that, no matter what parameter of the sys-
gration over the phase space, yields an HO Wehrl’s tem we choose to measure, I has to be larger or equal
entropy than the inverse of the mean-square error associated
 with the concomitant experiment. This result,
W (HO) = 1 − ln 1 − e−β h̄ω , (13)
I e2
1, (16)
that is the Lieb’s HO-thermal uncertainty relation [12].
Finally, notice that is referred to as the Cramer–Rao bound [2]. The cel-
ebrated Uncertainty Principle of Heisenberg’s can be

dp dx  shown to constitute a special instance of (16) [2]. On
2
e|z| (β, ω) ≡ e|z|
2
= µ(p , x  )|z|2
π account of (16) one is in a position to state that I pro-

 2 vides us with a positive amount of information [2],
dp dx   
− µ(p , x )|z| as opposite to Shannon’s entropy, that measures igno-
π rance [13]. Also, the latter is a global measure, while
  2
= |z|2 − |z| I is a local one [2]. If y1 , y2 , . . . , yn are n relevant pa-
 rameters of the problem at hand (possibly including θ ,
dp dx  
= µ(p , x  ) p 2 + x  2 but not, of course, x, that is integrated over), we will
π rewrite (16) in the fashion
1
= , (14) F (y1 , y2 , . . . , yn ) ≡ I e2
1. (17)
1 − e−β h̄ω
ranges equals unity (Heisenberg’s uncertainty lowest In the case of the harmonic oscillator, for instance,
limit) for T = 0 and diverges as T → ∞, the typical these parameters are the inverse temperature and the
behavior of a thermal uncertainty relation. frequency.
A particular I -case is of great importance: that
2.2. A brief primer on Fisher’s information measure of translation families [2,15], i.e., that in which I
is a functional of distribution functions (DF) whose
A very important information measure is that ad- form does not change under θ -displacements. These
vanced by R.A. Fisher in the twenties (a detailed study DF are shift-invariant (à la Mach, no absolute origin
can be found in Refs. [1,2]). Let us consider a system for θ ), and for them Fisher’s information measure
that is specified by a physical parameter θ , while x is a (FIM) adopts the somewhat simpler appearance [2]
  
stochastic variable (x ∈ N ) and fθ (x) the probability ∂ ln f (x) 2
density for x, which depends on the parameter θ . An I (shift-invariant) = dx f (x) . (18)
∂x
 F. Pennini, A. Plastino / Physics Letters A 326 (2004) 20–26 23

This shift-invariant form of I has encountered many so-called q-logarithmic function [6], a generalization
physical applications [2]. of the standard logarithmic function. The real parame-
We will be concerned below with a special FIM- ter q is called the index of nonextensivity, the conven-
form called the “escort Fisher measure”. It was de- tional Boltzmann–Gibbs statistics being recovered in
vised, for a nonextensive setting [6,7], by Plastino, the limit q → 1.
Plastino, Miller, and Pennini in [15,16]. Let us remind A typical feature of nonextensive thermostatistics is
the reader first of all of the useful concept of escort that of employing expectation values constructed with
probabilities (see [17] and references therein). Given escort PDs. This is, if the quantity ∆ takes the value
a normalized, discrete (continuous) probability distri- ∆i for the event i of probability P (i), then [15]
bution (PD) P (i) (f (x)), its associated escort PD of

order q (q any real parameter) is defined, for the dis- ∆q = P (q) (i)∆i , (23)
crete or continuous case as, respectively, [17], i
is to be regarded as the expectation value of ∆ in using
P (i)q
P (q) (i) =  q
, or the MaxEnt approach [13] in conjunction with Sq [21,
i P (i) 24]. Tsallis’ nonextensivity index is thereby identified
f (x)q with the order of the underlying escort distribution, as
f (q) (x) =  . (19)
dx f (x)q first pointed out in [15]. Summing up, current usage
of nonextensive thermostatistics employs three basic
In the case of complex scenarios involving a PD ingredients:
P (i) (f (x)), it is often the case that the associated
escort PDs yield more insights into the concomitant (1) Tsallis entropy;
dynamics than the original PD [17]. The escort FIM is (2) MaxEnt;
then just Fisher’s measure expressed as a functional of (3) q-expectation values evaluated with escort distri-
a escort distribution of order q [15] butions.
  (q) 
(q) ∂ ln fθ (x) 2
I (q) (θ ) = dx fθ (x) , (20) 2.4. Fisher measure and Husimi distributions
∂θ
which obeys, instead of (16), the “escort Cramer–Rao For the reader’s convenience, we summarize first
relation” [15] of all results obtained in [5] that involve the shift-
invariant Fisher measure associated to the Husimi
Fq ≡ q 2 I (q) eq2
1, (21) probability distribution µ(p, x). Firstly, remember
where eq2 , of course, stands for the mean-square error that Fisher’s measure is additive [2]: if x and p are
evaluated with the escort distribution (compare with independent variables, I (p + x) = I (p) + I (x), where
Eq. (14)). we denote for θ ≡ τ = (p, x) a point in phase-space,
so that we face a shift-invariance situation. One defines
2.3. Nonextensive thermostatistics and escort z in terms of the variables x and p, that are scaled
distributions by their respective variances (cf. above the definition
of |z). The ensuing shift-invariant Fisher measure is
Nonextensive thermostatistics is regarded by many then [5]

authors as a new paradigm for statistical mechanics dp dx
I (shift-invariant) ≡ Iτ = µ(p, x)Γ (24)
(see, for instance, [6,7,18–24] and references therein). 2π h̄
It is based on Tsallis’ nonextensive information mea- with
sure


 ∂ ln µ(p, x) 2 ∂ ln µ(p, x) 2
Γ = σx2 + σp2 , (25)
Sq = − dx p(x)q lnq p(x), (22) ∂x ∂p
so that we can recast it in the form
where p(x) is a normalized probability density defined  2
for x ∈ N and lnq (x) = (x 1−q − 1)/(1 − q) is the Γ = 1 − e−h̄βω |z|2 . (26)
24 F. Pennini, A. Plastino / Physics Letters A 326 (2004) 20–26

The above measure (24) constitutes an estimation In the limit q → 1 we have the standard form obtained
tool for location in phase-space. Using here the HO by Anderson et al. given by Eq. (13), since
Husimi µ-expression (5), Iτ adopts the appearance  −1  
lim lnq 1 − e−β h̄ω = − ln 1 − e−β h̄ω .
Iτ (HO) = 1 − e−β h̄ω , (27) q→1

Note that, when the temperature goes to zero (β →

so that, using (14), we immediately verify that
∞), then
F (β, ω) ≡ Iτ e|z|
2
= 1, (28)
W (q) (HO) → q. (32)
i.e., Cramer–Rao bound is reached. This result is liable
This provides us with a new interpretation for Tsallis’
to arouse mixed feelings. On the positive side, one sees
nonextensivity index q. It is the q-Wehrl entropy
that efficient estimation is possible at all temperatures,
of an HO at T = 0. Additionally, it follows that,
not only at T = 0. On the debit side, however, we
in a quantal regime, q cannot be negative. Indeed,
lose in F (β, ω) all temperature-dependence in our
according to the most basic tenet of information
à la Cramer Fisher estimation process. We will remedy
theory, W (q) represents our ignorance with regards
this situation below by recourse to escort distributions.
to location in phase-space once we know that the
Finally, notice also that comparison with Eq. (13)
probability distribution for τ is µ(τ ) [13]. Obviously,
allows us to write [5]
this ignorance-amount cannot be negative. Thus, we
 
W (HO) = 1 − ln Iτ (HO) obtain a physical lower-bound for q
⇒ W + ln[Iτ ] = 1, (29) q
0. (33)
and we regain contextual temperature information, but There is more, however. On account of the Lieb bound
using both Wehrl’s and Fisher’s measures. Since the W
1 [12], we also get
first one manages to do this by itself (cf. Eq. (13)), not
too much is gained. What we really want is to obtain q
1. (34)
such a temperature context by recourse to Fisher’s
information by itself, without further ado. 3.2. Escort Husimi distributions

It is now appropriate to introduce an escort q-Husi-

3. The Husimi–Tsallis distribution mi distribution in the fashion
µ(p, x)q
3.1. q-Wehrl measure γq (p, x) =  dp dx
. (35)
q
2π h̄ µ(p, x)
We will use in what follows the abbreviation τ ≡ The associated q-Husimi–Fisher measure (24) for
(p, x) and proceed now with the task of generaliz- translation families is then
ing Wehrl’s information measure (7) so as to accom-   dp dx q
modate it to a Tsallis nonextensive environment and dp dx 2π h̄ µ(p, x) Γ
Iτ =
(q)
γq (p, x)Γq = q  dp dx
2
,
thus obtain the concomitant “nonextensive Wehrl en- 2π h̄ q
2π h̄ µ(p, x)
tropy” W (q) . This is straightforwardly achieved in the (36)
fashion
 since Γq = q 2 Γ . Eq. (36) constitutes an escort Husimi
dp dx
W =−
(q)
µ(p, x)q lnq µ(p, x), (30) estimation tool for location in phase-space.
2π h̄
the integration process encompassing the whole of 3.3. HO application
phase-space. Explicit evaluation of (30) yields, for the
thermal HO, 3.3.1. The q-Wehrl–Fisher connection
  −1  Let us connect now the measure W (q) with the
W (q) (HO) = q 1 + lnq 1 − e−β h̄ω . (31) shift-invariant Fisher one Iτ through the HO Wehrl
 F. Pennini, A. Plastino / Physics Letters A 326 (2004) 20–26 25

generalization as we had previously ascertained following a

  Wehrl route.
W (q) (HO) = q 1 − Iτq−1 lnq Iτ , (37) • Comparing Eq. (43) with the information (27) we
where we have used the facts that (i) lnq (1/x) = find the following relation:
−x q−1 lnq x, ∀x, ∀q, and (ii) when the parameter q
tends to the unity W (1) ≡ W . In the HO-instance we Iτ(q) (HO) = qIτ (HO), (45)
have, (i)
 which connects the escort q-Fisher information
dp dx 1 q−1 for translation families with the original (q = 1)
µ(p, x)q = 1 − e−β h̄ω , (38)
2π h̄ q one.
so that we can check that, in the limit q → 1, the
function µ is normalized to unity, and (ii) 3.3.3. State equation for q
 It has been speculated in the literature that, in some
dp dx 1 q−2
µ(p, x)q |z|2 = 2 1 − e−β h̄ω . (39) instances, one could face a temperature-dependent
2π h̄ q q = q(T ) nonextensivity index (see, for instance, [24–
26], and references therein). In such a vein, let us
3.3.2. Excitation energy
assume that the parameter q is indeed a function of β.
It is opportune to recall at this point that |z|2 is
Consideration of the q-Cramer–Rao bound (21) leads
proportional to the excitation energy E [11]
to the following idea, in order to gain more insight into
h̄ω (q)
(44): extremize qIτ by deriving it with respect to β
E(z) = z|Ĥ |z − = h̄ω|z|2 = z|h̄ωâ † â|z. (40)
2 so as to obtain an equation for q as a function of the
Several items are here to be emphasized: inverse temperature. Setting

• From (1), (25), (36), and (40) (q)

d[qIτ ]
 dp dx = 0, (46)
dβ
 2π h̄ µ(p, x) |z|
q 2
Iτ(q) (HO) = q 2 1 − e−h̄βω
2
 dp dx we obtain a differential equation for q
q
2π h̄ µ(p, x)
 2
≡ |z| q , (41) dq  β h̄ω
e − 1 + q h̄ω = 0 (47)
dβ
i.e.,
 2  whose solution is of the form
Iτ(q) (HO) = q 2 1 − e−h̄βω E/h̄ω q , (42)
 −1
which relates the q-Fisher information measure to q(β) = 1 − e−β h̄ω , (48)
the γ -thermal mean value of the excitation energy
and thus we find an escort Cramer–Rao relation
(the q sub-index indicates that we are employing
q-mean values).  −1
Fq (β, ω) ≡ q = 1 − e−β h̄ω . (49)
• One finds, by inserting (38) and (39) into (36),
 The q-Cramer–Rao bound Fq (β, ω) becomes (cf.
Iτ(q) (HO) = q 1 − e−β h̄ω . (43)
Eq. (14)) a thermal uncertainty relation (TUR) as that
• Eq. (43) tells us, once again, that q cannot be of Lieb’s [12]. In such a sense one can then argue
negative, since it is the ratio of two positive- that, with q = q(β) one “optimizes” the information
definite quantities. measure in the sense of transforming the associated
• Since one easily verifies that eq2 in Eq. (21) bound into a TUR. The “equation of state” (47)
verifies eq2 (HO) = e|z|
2 (HO)/q 2 , the q-Cramer– expresses the nonextensivity index q in terms of the
Rao inequality (16) reads here (cf. Eq. (28)) temperature and the frequency. In particular, at zero
temperature we see that q = 1. The present is the first
Fq (β, ω) = q 2 Iτ(q) eq2 = qIτ e2 ≡ q
1 concrete example, as far as we know, of a temperature-
⇒ q
1, (44) dependent q = q(T ) nonextensivity index.
26 F. Pennini, A. Plastino / Physics Letters A 326 (2004) 20–26

4. Conclusions [2] B.R. Frieden, Physics from Fisher Information, Cambridge

Univ. Press, Cambridge, 1998.
In this work, by recourse to the concept of escort- [3] A. Wehrl, Rep. Math. Phys. 16 (1979) 353.
[4] A. Anderson, J.J. Halliwell, Phys. Rev. D 48 (1993) 2753.
distribution, we have performed a study of generalized [5] F. Pennini, A. Plastino, cond-mat/0312680.
information measures (GIM) constructed with the [6] M. Gell-Mann, C. Tsallis (Eds.), Nonextensive Entropy: Inter-
Husimi distributions. Investigating the connection of disciplinary Applications, Oxford Univ. Press, Oxford, 2004.
these GIMs with (i) the Wehrl entropy and (ii) thermal [7] A. Plastino, A.R. Plastino, Braz. J. Phys. 29 (1999) 50.
uncertainty relations, has allowed us to find: [8] V. Latora, A. Rapisarda, C. Tsallis, Physica A 305 (2002) 129,
and references therein.
[9] J.R. Klauder, B.S. Skagerstam, Coherent States, World Scien-
• Two new interpretations for the nonextensivity tific, Singapore, 1985.
index q in terms of [10] K. Husimi, Proc. Phys. Math. Soc. Jpn. 22 (1940) 264.
(1) the q-Wehrl entropy, and [11] J. Schnack, Europhys. Lett. 45 (1999) 647.
(2) the q-escort Cramer–Rao bound, [12] E.H. Lieb, Commun. Math. Phys. 62 (1978) 35.
[13] A. Katz, Principles of Statistical Mechanics, Freeman, San
for the case of the thermal quantum harmonic Francisco, CA, 1967.
oscillator. [14] H. Cramer, Mathematical Methods of Statistics, Princeton
• A lower bound for q, namely, q = 1, obtained in Univ. Press, Princeton, NJ, 1946.
two different ways. [15] F. Pennini, A.R. Plastino, A. Plastino, Physica A 258 (1998)
• A relation between the q-Fisher information mea- 446.
[16] A. Plastino, A.R. Plastino, H.G. Miller, Physica A 235 (1996)
sure and the thermal mean value of the excitation 577.
energy, valid also for q = 1. [17] C. Beck, F. Schlögl, Thermodynamics of Chaotic Systems,
• A “state-equation” that gives q as a function Cambridge Univ. Press, Cambridge, 1993.
of T , ω and transforms the q-Cramer–Rao bound [18] C. Tsallis, Chaos Solitons Fractals 6 (1995) 539;
(q) C. Tsallis, Phys. World 10 (1997) 42.
[15] Fq (β, ω) = q 2 Iτ eq2
1 into a thermal [19] C. Tsallis, Braz. J. Phys. 29 (1999) 1, and references therein.
uncertainty relation. [20] A.R. Plastino, A. Plastino, Phys. Lett. A 193 (1994) 140.
[21] C. Tsallis, R.S. Mendes, A.R. Plastino, Physica A 261 (1998)
The HO is, of course, much more than a mere 534.
example. Nowadays it is of particular interest for the [22] S. Martínez, F. Pennini, A. Plastino, Physica A 295 (2001) 416.
[23] S. Abe, S. Martínez, F. Pennini, A. Plastino, Phys. Lett. A 281
dynamics of bosonic or fermionic atoms contained in (2001) 126.
magnetic traps [27–29] as well as for any system that [24] S. Martínez, F. Nicolás, F. Pennini, A. Plastino, Physica A 286
exhibits an equidistant level spacing in the vicinity of (2000) 489.
the ground state, like nuclei or Luttinger liquids. [25] M. Casas, S. Martínez, F. Pennini, A. Plastino, Physica A 305
(2002) 41.
[26] S. Abe, Physica A 269 (1999) 403.
[27] M.H. Anderson, et al., Science 269 (1995) 198.
References [28] K.B. Davis, et al., Phys. Rev. Lett. 75 (1995) 3969.
[29] C.C. Bradley, C.A. Sackett, R.G. Hulet, Phys. Rev. Lett. 78
[1] B.R. Frieden, B.H. Soffer, Phys. Rev. E 52 (1995) 2274. (1997) 985.