Astrophys Space Sci (2006) 305:241–245 DOI 10.1007/s10509-006-9198-5 O R I G I NA L A RT I C L E Why q-Expectation Values Must be Used in Nonextensive Statistical Mechanics Sumiyoshi Abe Received: 2 May 2006 / Accepted: 2 June 2006 C Springer Science + Business Media B.V. 2006 Abstract There is a controversy in the area of nonextensive statistical mechanics regarding the form of the expectation value of a physical quantity. Two definitions have been discussed in the literature: one is the ordinary definition and the other is the normalized q-expectation value associated with the escort distribution. Here, it is proved that the normalized q-expectation value is the correct one to be employed. The Shore-Johnson theorem is used to show that the formalism with the normalized q-expectation value is theoretically consistent with minimum cross entropy principle, whereas the ordinary expectation value has to be excluded. Keywords q-expectation values . nonextensive statistical mechanics 1. Introduction Nonextensive statistical mechanics (Abe and Okamoto, 2001; Kaniadakis et al., 2002; Gell-Mann and Tsallis, 2004; Kaniadakis and Lissia, 2004) pioneered by Tsallis (1988) has been attracting continuous interest over the years. It is expected to offer a unified framework for describing complex systems in their nonequilibrium stationary states, systems with (multi)fractal and self-similar structures, longrange interacting systems, anomalous diffusion phenomena, and so on. The worked examples include cosmic rays (Tsallis et al., 2003), astrophysics and self-gravitating systems (Lavagno et al., 1998; Taruya and Sakagami, 2003a,b), dynamical systems at the edge of chaos (Latora et al., 2000; S. Abe Institute of Physics, University of Tsukuba, Ibaraki 305-8571, Japan e-mail: [email protected] de Moura et al., 2000; Baldovin and Robledo, 2002; Johal and Tirnakli, 2004), lattice Boltzmann models (Boghosian et al., 2003), magnetism of colossal magnetoresistance manganites (Reis et al., 2002), high-energy processes (Walton and Rafelski, 2000; Bediaga et al., 2000; Beck, 2000; Alberico et al., 2000; Navarra et al., 2003), quantum groups (Abe, 1997, 1998, 2003b), cellular aggregates (Upadhyaya et al., 2001), Lévy flights (Prato and Tsallis, 1999; Abe and Rajagopal, 2000a), semiclassical dynamics in optical lattices (Lutz, 2004), kinetics of charged particles (Rossani and Scarfone, 2000), Internet traffic (Abe and Sukuki, 2003a), earthquakes (Abe and Suzuki, 2003a, 2005; SotolongoCosta and Posadas, 2004), econophysical problems (Borland, 2002a,b; Kozuki and Fuchikami, 2003), and complex networks (Tadic and Thurner, 2004; Abe and Sukuki, 2004a; Wilk and Wlodarczyk, 2004). In spite of these successes, the theoretical foundations of nonextensive statistical mechanics are yet to be established in some respects. Among others, the problem concerning the definition of expectation value of a physical quantity seems to be poorly understood. There are two different definitions in the literature. One is the frequently employed definition, which is the normalized q-expectation value (Tsallis et al., 1998): U (nor) = H q = Pi εi , (1) ( pi )q Pi = , q j(pj) (2) i where Pi is called the escort distribution (Beck and Schlögl, 1993; Abe, 2003c) associated with the basic distribution pi and H is a physical random variable under consideration (e.g., the system energy) with its ith value εi .q is taken to be Springer 242 Astrophys Space Sci (2006) 305:241–245 positive. The other definition is the ordinary one: U (ord) = H = −β pi εi , where α and β are the Lagrange multipliers. After eliminating α, the resulting maximum entropy distribution p̃i(ord) is found to be 1/(q−1) p̃i(ord) = 1 + (1 − q) S̃q(ord) 1/(q−1) q −1 × 1− , β εi − Ũ (ord) q + (6) where β = i β p̃i(ord) q , (7) and [a]+ ≡ max{0, a}. In Equation (6), S̃q(ord) and Ũ (ord) are calculated in terms of the distribution in Eq. (6) in the selfreferential manner. On the other hand, if the normalized q-expectation value is employed, the functional to be maximized reads (nor) [ p; α, β] = Sq [ p] − α −β 1 p̃i(nor) = First let us recall how nonextensive statistical mechanics depends on the definition of expectation value. In nonextensive statistical mechanics, the Tsallis entropy indexed by q (Tsallis, 1988) pi − 1 i ( pi )q εi i q j(pj) −U (nor) . (8) Z̃ q(nor) 1/(1−q) 1 − (1 − q)β ∗ εi − Ũ (nor) + , 1/(1−q) Z̃ q(nor) = 1 + (1 − q) S̃q(nor) 1/(1−q) 1 − (1 − q)β ∗ εi − Ũ (nor) + = , (9) (10) i kB q sq [ p] = ( pi ) − 1 1−q i (4) is maximized under appropriate constraints on the normalization condition of the probability distribution and the expectation value of a physical quantity under consideration. Here and hereafter, the Boltzmann constant, k B , is set equal to unity for simplicity. If the ordinary expectation value is employed, then the functional to be maximized is i Springer (5) The corresponding maximum entropy distribution is given by 2. Ordinary Expectation Value and Normalized q-Expectation Value [ p; α, β] = Sq [ p] − α pi εi − U (ord) , (3) which is preferred by some researchers. This may be due to the fact that the nonextensive statistical mechanical formalisms with these two definitions of expectation value lead to the stationary distribution of a similar kind (see Sec. 2). Therefore, it is crucial to identify which the correct definition is. It has been shown (Abe and Rajagopal, 2000) that, for a class of power-law distributions (i.e., the q-exponential distribution with q > 1), only the normalized q-expectation value is consistent with the method of steepest descents for deriving canonical ensemble from microcanonical ensemble. The situation, however, remains unclear for the other class of distributions with compact supports (i.e., the q-exponential distribution with 0 < q < 1). In this paper, we show that the definition to be employed in nonextensive statistical mechanics is the normalized qexpectation value. For this purpose, we study the properties of two generalized relative entropies regarding the two different definitions of expectation value. Then, we use the Shore-Johnson theorem for minimum cross entropy (relative entropy) principle to prove that the formalism with the normalized q-expectation value is theoretically consistent, whereas the ordinary expectation value cannot be employed. i i (ord) pi − 1 where β∗ = i β p̃i(nor) q . (11) Similarly to the case of the ordinary expectation value, S̃q(nor) and Ũ (nor) are the values of Sq and U (nor) calculated in terms of the maximum entropy distribution p̃i(nor) in the self-referential manner. Clearly, the Tsallis entropy in Equation (4) and the normalized q-expectation value tend to the Boltzmann-GibbsShannon entropy S[ p] = − i pi ln pi and the ordinary expectation value in the limit q → 1. Accordingly, both of Astrophys Space Sci (2006) 305:241–245 243 the distributions in Equations (6) and (9) converge to the Boltzmann-Gibbs distribution p̃i ∼ exp(−βεi ) in such a limiting case. It is mentioned that they resemble in their forms (with opposite signs of the exponents). This point may be a reason why some researchers think that there is no essential reason to prefer the normalized q-expectation value and the ordinary expectation value may be used. A point to be noticed is that in both cases the following thermodynamic relations hold: ∂ S̃q(ord) ∂ Ũ (ord) ∂ S̃q(nor) ∂ Ũ (nor) = β, (12) = β, (13) which guarantee the existence of the thermodynamic Legendre-transform structure in both cases. However, it is still an open problem in nonextensive statistical mechanics what the physical temperature is (Abe, 2001; Abe et al., 2001). 3. Generalized relative entropies associated with ordinary expectation value and normalized q-expectation value Relative entropy has an important physical meaning as free energy difference. It is known in mathematical information theory that there exist two different kinds of relative entropies: one is of the Bregman type (Bregman, 1967) and the other the Csiszár type (Csiszár, 1972). The Bregman-type relative entropy (Naudts, 2004) and the Csiszár-type relative entropy (Abe, 1998; Tsallis, 1998b; Borland et al., 1998, 1999) associated with the Tsallis entropy are respectively given by Iq [ p r ] = 1 pi [( pi )q−1 − (ri )q−1 ] q −1 i − ( pi − ri )(ri )q−1 , i K q [ p r ] = (14) i pi ln pi . ri pi d s[s q−1 − (ri )q−1 ], (17) ri we immediately see its nonnegativity. For K q [ p r ], it is convenient to rewrite it as K q [ p r ] = 1 pi [1 − (ri / pi )1−q ]. 1−q i (18) Then, using the inequality, (1 − x 1−q )/(1 − q) ≥ 1 − x for x > 0 and q > 0 with the equality for x = 1, we also see nonnegativity of K q [ p r ]. In what follows, we discuss the physical meanings of Iq [ p r ] and K q [ p r ]. Let us take the maximum entropy distributions as the reference distributions in Iq [ p r ] and K q [ p r ]. Taking into account the exponents of the maximum entropy distributions in Equations (6) and (9) together with the dependencies of Iq [ p r ] and K q [ p r ] on ri , we can expect that Iq [ p r ] and K q [ p r ] may be associated with the formalisms with the ordinary expectation value and the normalized q-expectation value, respectively. This is indeed the case, as we shall see below. Substituting ri = p̃i(ord) into Equation (14), we have Iq [ p p̃ (ord) ] = β Fq(ord) − F̃q(ord) , (19) where Fq(ord) = U (ord) − 1 Sq , β F̃q(ord) = Ũ (ord) − 1 (ord) S̃ . β q (20) On the other hand, putting ri = p̃i(nor) into Equation (15), we have K q [ p p̃ (nor) ] = β̂ q p̃i(nor) Fq(nor) − F̃q(nor) , (21) where (15) β̂ = β ∗ ( pi )q , (22) i where ri is a reference distribution (i.e., prior). In the limit q → 1, both Iq [ p r ] and K q [ p r ] tend to the KullbackLeibler relative entropy H [ p r ] = q q −1 i Iq [ p r ] = i 1 1− ( pi )q (ri )1−q , 1−q i Like H [ p r ], Iq [ p r ] and K q [ p r ] are nonnegative and vanish if and only if pi = ri (∀i). This can be seen as follows. Noticing that Iq [ p r ] admits the integral representation (Naudts, 2004) (16) Fq(nor) = U (nor) − 1 Sq , β̂ F̃a(nor) = Ũ (nor) − 1 (nor). S̃q β̂ (23) Equations (19) and (21) imply that Iq [ p r ] and K q [ p r ], in fact, give the “free energy differences” and, therefore, are identified with the generalized relative entropies associated with the ordinary expectation value and the normalized Springer 244 Astrophys Space Sci (2006) 305:241–245 q-expectation value, respectively. We also mention that the quantum mechanical counterpart of K q [ p r ] has recently been employed to prove the second law of thermodynamics (Abe and Rajagopal, 2003). Now, convexity is one of the most important properties to be satisfied by any types of relative entropies. As can be seen, Iq [ p r ] is convex in pi , but not in ri . In marked contrast to this flaw, K q [ p r ] is, like the Kullback-Leibler relative entropy in Equation (16), jointly convex (Abe, 2003a, 2004). Kq a λa p(a) λa r(a) ≤ λa K q [ p(a) r(a) ], a a (24) where λa > 0 and a λa = 1. Finally, we mention that, like the Kullback-Leibler relative entropy, K q [ p r ] is “composable” (Tsallis, 2001), but Iq [ p r ] is not. For factorized joint distributions of a bipartite system (A, B), pi j (A, B) = p(1)i (A) p(2) j (B) and ri j (A, B) = r(1)i (A)r(2) j (B), K q [ p(1) P(2) r(1)r(2) ] satisfies the following relation: K q [ p(1) p(2) r(1)r(2) ] = K q [ p(1) r(1) ] + K q [ p(2) r(2) ] +(q − 1)K q [ p(1) r(1) ]K q [ p(2) r(2) ], (25) whereas such a closed relation does not exist for Iq [ p(1) p(2) r(1)r(2) ]. Axiom III (System independence): It should not matter whether one accounts for independent information about independent systems separately in terms of their marginal distributions or in terms of the joint distribution. Axiom IV (Subset independence): It should not matter whether one treats independent subsets of the states of the systems in terms of their separate conditional distributions or in terms of the joint distribution. Axiom V (Expansibility): In the absence of new information, the prior (i.e., the reference distribution) should not be changed. These axioms are natural in the sense that all of them are fulfilled by the ordinary Kullback-Leibler relative entropy in Equation (16), which gives rise to the free energy difference in Boltzmann-Gibbs statistical mechanics. For the Tsallis entropy in Equation (4), the axioms and uniqueness theorem are known in the literature (dos Santos, 1997; Abe, 2000). In contrast to this fact, the above set of axioms is quite general and not very restrictive. Accordingly, they do not uniquely specify the explicit functional form of the relative entropy. Now, the Shore-Johnson theorem (Shore and Johnson 1980, 1981, 1983) states that the relative entropy J [ p||r ] with the prior ri and the posterior pi satisfying the axioms I–V must have the following form: J [ p||r ] = pi h( pi /ri ), (26) i 4. Shore-Johnson theorem selects normalized q-expectation value In this section, we shall see by using the Shore-Johnson theorem that K q [ p r ] associated with the normalized qexpectation value leads to a consistent framework for minimum cross entropy (i.e., relative entropy) principle, whereas Iq [ p r ] corresponding to the ordinary expectation value does not. About a quarter a century ago, Shore and Johnson (1980, 1981, 1983) have presented a set of axioms for minimum cross entropy (i.e., relative entropy) principle. They have made an attempt to answer to the question: why the correct rule of inference is to minimize relative entropy, in conformity with a vindication of Jaynes’ claim that every other rule will lead to contradiction (Uffink, 1995). The five key axioms are listed as follows. Axiom I (Uniqueness): If the same problem is solved twice, then the same answer is expected to result both times. Axiom II (Invariance): The same answer is expected when the same problem is solved in two different coordinate systems, in which the posteriors in the two systems should be related by the coordinate transformation. Springer where h(x) is some function. It is crucial to recognize that the function h(x) surely exists for K q [ p||r ]: h(x) = 1 (1 − x q−1 ) . 1−q (27) On the other hand, Iq [ p||r ] cannot be recast to the form in Equation (26), since Iq [ p||r ] may violate Axiom III. Therefore, we conclude that the Shore-Johnson theorem supports the normalized q-expectation value and excludes the possibility of using the ordinary expectation value from nonextensive statistical mechanics. 5. Concluding remarks We have studied the meanings of the two different definitions of expectation value in nonextensive statistical mechanics, i.e., the ordinary expectation value and the normalized q-expectation value. To identify which the correct definition is, we have discussed the corresponding two kinds of generalized relative entropies, which have the physical meanings as the “free energy differences”. It was shown that Astrophys Space Sci (2006) 305:241–245 the Shore-Johnson theorem for minimum cross entropy (i.e., relative entropy) principle supports the generalized relative entropy associated with the formalism with the normalized q-expectation value and excludes the possibility of using the ordinary expectation value from nonextensive statistical mechanics. Thus, we conclude that what to be employed in nonextensive statistical mechanics is not the ordinary expectation value but the normalized q-expectation value. Acknowledgements The author thanks Hans J. Haubold for inviting him to the Twelfth United Nations/European Space Agency Workshop on Basic Space Science (24–28 May 2004, Beijing, P. R. China). This work was supported in part by the Grant-in-Aid for Scientific Research of Japan Society for the Promotion of Science. References Abe, S., Okamoto, Y. (eds): Nonextensive Statistical Mechanics and Its Applications. Springer-Verlag, Heidelberg (2001) Abe, S.: Phys. Lett. A 312, 336 (2003a) Abe, S.: Phys. Lett. A 324, 507 (2004) (E). See also Abe, S.: Physica A 344, 359 (2004) Abe, S., Rajagopal, A.K.: J. Phys. A 33, 8723 (2000a) Abe, S., Rajagopal, A.K.: J. Phys. A 33, 8733 (2000) Abe, S., Rajagopal, A.K.: Phys. Rev. Lett. 91, 120601 (2003) Abe, S., Sukuki, N.: Europhys. Lett. 65, 581 (2004) Abe, S., Suzuki, N.: J. Geophys. Res. 108(B2), 2113 (2003a) Abe, S., Suzuki, N.: Phys. Rev. E 67, 016106 (2003a) Abe, S., Suzuki, N.: Physica A 350, 588 (2005) Abe, S.: in Proceedings of the 5th International Wigner Symposium, Vienna, 1997, edited by Kasperkovitz, P., Grau, D. (eds.) (World Scientific, Singapore, 1998), p. 66. See also Rajagopal, A.K., Abe, S.: Phys. Rev. Lett. 83, 1711 (1999) Abe, S.: Phys. Lett. A 224, 326 (1997) Abe, S.: Phys. Lett. A 244, 229 (1998) Abe, S.: Phys. Lett. A 271, 74 (2000) Abe, S.: Phys. Rev. A 68, 032302 (2003) Abe, S.: Phys. Rev. E 63, 061105 (2001). See also Abe, S., Martı́nez, S., Pennini, F., Plastino, A.: Phys. Lett. A 281, 126 (2001) Abe, S.: Phys. Rev. E 68, 031101 (2003c) Abe, S.: Physica A 300, 417 (2001) Alberico, W.M., Lavagno, A., Quarati, P.: Eur. Phys. J. C 12, 499 (2000) Baldovin, F., Robledo, A.: Europhys. Lett. 60, 518 (2002a) Baldovin, F., Robledo, A.: Phys. Rev. E 66, 045104 (2002b) Beck, C., Schlögl, F.: Thermodynamics of Chaotic Systems: An Introduction, Cambridge University Press, Cambridge (1993) Beck, C.: Physica A 286, 164 (2000) Bediaga, I., Curado, E.M.F., de Miranda, J.M.: Physica A 286, 156 (2000) 245 Boghosian, B.M., Love, P.J., Coveney, P.V., Karlin, I.V., Succi, S., Yepez, J.: Phys. Rev. E 68, 025103 (2003) Borland, L., Plastino, A.R., Tsallis, C.: J. Math. Phys. 39, 6490 (1998) Borland, L., Plastino, A.R., Tsallis, C.: J. Math. Phys. (E) 40, 2196 (1999) Borland, L.: Phys. Rev. Lett. 89, 098701 (2002a) Borland, L.: Quantitative Finance 2, 415 (2002b) Bregman, L.M.: USSR Comp. Math. Math. Phys. 7, 200 (1967) Csiszár, I.: Per. Math. Hung. 2, 191 (1972) de Moura, F.A.B.F., Tirnakli, U., Lyra, M.L.: Phys. Rev. E 62, 6361 (2000) dos Santos, R.J.V.: J. Math. Phys. 38, 4104 (1997) Gell-Mann, M., Tsallis, C. (eds.): Nonextensive Entropy– Interdisciplinary Applications, Oxford University Press, New York (2004) Johal, R.S., Tirnakli, U.: Physica A 331, 487 (2004) Kaniadakis, G., Lissia, M. (eds.): Special Issue of Physica A 340 (2004) Kaniadakis, G., Lissia, M., Rapisarda, A. (eds.): Special Issue of Physica A 305 (2002) Kozuki, N., Fuchikami, N.: Physica A 329, 222 (2003) Latora, V., Baranger, M., Rapisarda, A., Tsallis, C.: Phys. Lett. A 273, 97 (2000) Lavagno, A., Kaniadakis, G., Rego-Monteiro, M., Quarati, P., Tsallis, C.: Astrophysical Letters and Communications 35, 449 (1998) Lutz, E.: Phys. Rev. A 67, 051402 (2003). See also Abe, S.: Phys. Rev. E 69, 016102 (2004) Naudts, J.: Rev. Math. Phys. 6, 809 (2004) Navarra, F.S., Utyuzh, O.V., Wilk, G., Wlodarczyk, Z.: Phys. Rev. D 67, 114002 (2003) Prato, D., Tsallis, C.: Phys. Rev. E 60, 2398 (1999) Reis, M.S., Freitas, J.C.C., Orlando, M.T.D., Lenzi, E.K., Oliveira, I.S.: Europhys. Lett. 58, 42 (2002) Rossani, A., Scarfone, A.M.: Physica A 282, 212 (2000) Shore, J.E., Johnson, R.W.: IEEE Transactions on Information Theory IT-26, 26 (1980) Shore, J.E., Johnson, R.W.: IEEE Transactions on Information Theory IT-27, 472 (1981) Shore, J.E., Johnson, R.W.: IEEE Transactions on Information Theory IT-29, 942 (1983) Sotolongo-Costa, O., Posadas, A.: Phys. Rev. Lett. 92, 048501 (2004) Tadic, B., Thurner, S.: Physica A 332, 566 (2004) Taruya, A., Sakagami, M.: Phys. Rev. Lett. 90, 181101 (2003a) Taruya, A., Sakagami, M.: Physica A 322, 285 (2003b) Tsallis, C., Anjos, J.C., Borges, E.P.: Phys. Lett. A 310, 372 (2003) Tsallis, C., Mendes, R.S., Plastino, A.R.: Physica A 261, 534 (1998) Tsallis, C.: in Ref. Kaniadakis et al. (2002). See also Abe, S.: Phys. Rev. E 63, 061105 (2001) Tsallis, C.: J. Stat. Phys. 52, 479 (1988a) Tsallis, C.: Phys. Rev. E 58, 1442 (1998b) Uffink, J.: Studies in History and Philosophy of Modern Physics 26B, 223 (1995) Upadhyaya, A., Rieu, J.P., Glazier, J.A., Sawada, Y.: Physica A 293, 549 (2001) Walton, D.B., Rafelski, J.: Phys. Rev. Lett. 84, 31 (2000) Wilk, G., Wlodarczyk, Z.: Acta Physica Polonica B 35, 871 (2004) Springer
© Copyright 2026 Paperzz