Thermodynamics of Energy Conversion and Transport
edited by S. Sieniutycz and A. De Vos
Springer, Berlin, 1999
From Statistical Distances to Minimally Dissipative
Processes
Chapter 11 by Lajos Diósi and Peter Salamon
A quantitative notion of statistical distinguishability led R.A. Fisher
to his idea of statistical distance which has since been developped into
Riemannian geometries on the space of statistical ensembles. Parallel to
though independently of this progress, Riemannian geometries were being
proposed on spaces of quantum states and also of thermodynamic states.
Riemannian geometries in various fields have found various applications
as different as in population dynamics and fractional distillation, just
to mention the first and the most recent ones. For decades, however, little
attention was being paid to the common theoretical basis of these geometric
methods. This Chapter intends to fill the gap. We present an elementary
introduction to the concept and mathematics of statistical distance in
order to help understand the emergence of Riemannian-geometrical structures.
While we put more emphasis on the thermodynamic aspects, the main goal
is still the interpretation of different applications on an equal footing
and using a unified framework.
1. Introduction
The Riemannian metric structure of thermodynamic theory, initiated
by Weinhold (1975) and Ruppeiner (1979), contains important and hitherto
barely tapped information concerning a physical system. The structure runs
deep; its presence can be felt at all levels of physical description. The
Riemannian metric of thermodynamics is, as shown first by Diósi
et al. (1984), in fact a realization of R. A. Fisher's concept of
statistical distinguishability (1922). He had applied it in 1922 to measure
genetic drift and later it became the basis for the mathematical theory
of information geometry. The corresponding notion of statistical distance
has since been introduced for various statistical systems. At the quantum
level the distance measures the reliability of experiments designed to
optimally distinguish between the two states along a one-parameter family
of density operators (Braunstein and Caves 1994). At the statistical mechanical
level, distance is the number of statistically distinguishable intermediate
states as we transform one state into another (Wootters 1981). This leads
to a natural Riemannian metric on the space of distributions in the thermodynamic
limit of Gibbs statistical ensembles. Numerous authors have speculated
about the meaning of the curvature defined by this geometry as a measure
of stability or interaction strength (c.f. Ruppeiner's recent review 1995).
The requirement of covariance with respect to this geometry can be used
to give an important correction to thermodynamic fluctuation theory. Finally,
at the macroscopic level, the square of this same distance between two
equilibrium states of a thermodynamic system equals the minimum entropy
produced in a process that transforms one state into the other, multiplied
by the number of relaxations during the transformation (Salamon and Berry
1983, Nulton et al. 1985). This result has become known as the horse-carrot
theorem. In this Chapter, we recapitulate basic ideas and results concerning
the Riemannian metric structure of thermodynamics while we attempt to shed
light on the underlying concept of statistical distance used in a much
broader context.
2. Empirical
Statistical Distance
The class of continuous variables spans from typical continuous quantities
of physics to approximate continuous quantities e.g. in population statistics.
Consider a continuous variable corresponding to a measurement of a real
number x to a certain precision Dx. The true value of x
lies in the confidence interval (x-Dx, x+Dx) with a probability
which amounts to 68% when Dx is, as we generally assume,
the standard deviation of x. The values Dx provide a measure
of distinguishability between different values of x. Two
values, say x and x', are statistically indistinguishable
if |x-x'| < Dx + Dx'. In the opposite case they are well-distinguishable.
By convention, we shall say that x and x' are statistically distinguishable
if the equality holds:
|x-x'| = Dx + Dx'.
2.1 Optimum calibration, 2.2 Naive optimum control,
2.3 More parameters
3. Theory
of Statistical Distance
The concept of statistical distinguishability, considered on intuitive
grounds in the previous section, implies a natural geometry on the space
of statistical ensembles. In the present section we give an insight into
the general structure of this geometry for classical as well as quantum
ensembles.
3.1 Classical Statistics, 3.2 Quantum Statistics
4. Riemannian
Geometry
We have shown in the previous section that the natural geometry, reflecting
statistical distinguishability of ensembles (probability distributions)
is non-Euclidean. In various fields of applications, we are dealing with
certain subclasses of probability distributions. Here we restrict ourselves
to the case where the probability distributions are parameterized by a
finite number of parameters. Gibbs distributions in statistical physics
are particularly relevant: they underly the statistical geometry of thermodynamic
parameter space.
4.1 Parameterized Statistics, 4.2 From Gibbs Statistics to Thermodynamics
5. Relevance
of Riemannian Geometry in Macroscopic Thermodynamics
We begin our treatment of macroscopic applications of the Riemannian
structure with a discussion of fluctuations since this follows most closely
from the arguments in the previous section. We will then turn our attention
to dissipation and a discussion of several horse-carrot theorems which
relate the dissipation associated with coaxing a system to traverse a given
sequence of states. The applications depend on the second order expansion
of the entropy and thus on the identity between the metric and the second
derivative of entropy.
5.1 A Covariant Fluctuation Theory, 5.2 Entropy Production, 5.3
The Metric as a Symmetric Product, 5.4 The group of transformations, 5.5
Dissipation in a small equilibration, 5.6 The discrete horse-carrot theorem,
5.7 Some comments on loss of availability, 5.8 The continuous horse-carrot
theorem, 5.9 A simple lemma from optimization, 5.10 Applications of the
lemma, 5.11 Cooling rates for simulated annealing
6. Staged
steady flow processes
Most recently (Salamon and Nulton 1998), the connection between dissipation
and geometry has been extended to treat a staged steady-flow process of
considerable industrial interest: fractional distillation. The example
involves some surprises which hint at the existance of other applications
of horse-carrot type analyses. The first surprise is that the scale of
the process is set by the flow rates rather than the states along the process.
The second surprise is that the null directions for our semi-Riemannian
metric turn out to be useful.
6.1 Dissipation in a Distillation Column
7. Conclusions
F. Weinhold (1975): ``Metric geometry of thermodynamics I-IV'', J.
Chem. Phys. 63 2479, 2484, 2488, 2496.
G.Ruppeiner (1979): ``Thermodynamics: a Riemannian geometric model'',
Phys. Rev. A 20 1608-1613.
L.Diósi, G.Forgács, B.Lukács and H.L.Frisch (1984):
``Metricization of thermodynamic state space and the renormalization group'',
Phys. Rev. A 29 3343-3345.
R.A. Fisher (1922): "On the dominance ratio", Proc. R. Soc. Edinburgh
42 321.
S.L. Braunstein and C.M. Caves (1994): ``Statistical Distance and the
geometry of quantum states'', Phys. Rev. Lett. 72 3439-3443.
W.K.Wootters (1981): ``Statistical distance and Hilbert space'', Phys.
Rev. D 23 357-362.
G.Ruppeiner (1995): ``Riemannian geometry in thermodynamic fluctuation
theory'', Rev. Mod. Phys. 67 605-659.
P. Salamon and R.S. Berry (1983): ``Thermodynamic Length and Dissipated
Availability'', Phys. Rev. Lett. 51 1127-1130.
J. Nulton, P. Salamon, B. Andresen, and Qi Anmin (1985): ``Quasistatic
Processes as Step Equilibrations'', J. Chem. Phys. 83 334.
P. Salamon and J.D. Nulton (1998): "The Geometry of Separation Processes:
The Horse-Carrot Theorem for Steady Flow Systems", Europhys. Lett. 42
571-576.