Saturday, March 06, 2010

Survey of Utility Theory?

1.0 Introduction
I think utility theory has a canonical textbook presentation. Many variations seem to exist. In some, the additional structure is imposed on the (commodity?) space over which agents choose. In others, more basic assumptions are made from which preferences can be derived under certain special cases.

I'd like to know if there are any surveys to read over these variations. I'm not insisting on something critical. And, given the dryness of the subject matter, I might not put such a survey on top of my queue. As can be seen below, I'm not sure of the field that would be demarcated by such a surveys. But literature surveys, in some sense, construct their object.

2.0 Textbook Treatment
Consider a space of n commodities. Each element of the space is a vector x = (x1, x2, ..., xn). Under the usual interpretation, xi is the quantity of the ith commodity.

An agent is modeled as having a preference relation, ≤, over the space of commodities. A typical question is what assumptions must hold for a utility function to exist. A utility function u(x) exists if, for all x and y in the space of commodities:
xy if and only if u(x) ≤ u(y)

Typically, the preference relation is taken to be a total order, that is, complete, reflexive, and transitive. A preference relation is complete if, for all x and y in the space of commodities,
xy or yx
A preference relation is reflexive if, for all x in the space of commodities
A preference relation is transitive if, for all x, y, and z in the space of commodities,
if xy and yz then xz

If the quantities of commodities fall along a continuum, a preference relation being a total order is not sufficient for a utility function to exist. Lexicographic preferences are an example of a preference relation for which a utility function does not exist. A continuity assumption rules out this case. This assumption is that for all x in the space of commodities, the sets {y | yx} and {z | xz} of commodities not preferred to x and commodities x is not preferred to, respectively, are closed.

Theorem: If a preference relation is a total order and is continuous in the above sense, then a utility function exists.

The utility function is only defined up to a monotonically increasing transformation. In other words, utility is ordinal. Typical exercises are to show certain properties of utility functions, such as ratios of marginal utilities (du/dxi)/(du/dxj), are invariant over the set of such transformations.

3.0 Probability
Von Neumann and Morgenstern generalized the commodity space to include vectors of the form: (p1, x(1); p2, x(2); ..., pm, x(m)), where:
p1 + p2 + ... + pm = 1
A commodity, in this sense, is a lottery. Each superscripted commodity vector x(i) is associated with a probability pi that it will be chosen.

Von Neumann and Morgenstern defined a new set of axioms to go along with their redefined commodity space. One implication is that for any two elements x and y in the commodity space, the linear combination (p, x; (1 - p), y) is also in the space. They obtain that a utility function exists, and it acts like mathematical expectation:
u(p1, x(1); p2, x(2); ..., pm, x(m)) = p1 u(x(1)) + p2 u(x(2)) + ... + pm u(x(m))

Under Von Neumann and Morgenstern's approach, utility functions are only defined up to affine transformations. That is, they are cardinal. In other words, they attain an interval measurement scale level. The utility for a lottery depends only on the probabilities and the resulting outcomes. It does not depend on how many spins of the wheel or roll of the dice are needed to decide between otherwise equivalent lotteries. Gambling is assumed to have no utility or disutility.

Leonard Savage develops axioms of probability concurrently with axioms of utility theory in his personalistic approach to probability and statistics. I'm not sure how much the survey I would like would go into approaches to probability, even if probability is important to decision theory. The same comment applies to game theory.

4.0 Attributes and Needs
Some see commodities as being chosen as an indirect means to choose something more abstract. As I understand it, Kevin Lancaster depicts a commodity as a bundle of attributes. Different commodities can have some attributes in common. A choice of an element in the space of commodities can then be related to an element in a space of commodity attributes.

The early Austrian school economists thought of goods as being desired for the satisfactions of wants. Water, for example, can be used to water your lawn, to satisfy a pet's thirst, or to drink yourself. One can imagine ranking wants in disparate categories. I am thinking of the triangular tables in Chapter III of Carl Menger's Principles of Economics, in Book III, Part A, Chapter III of Eugen von Böhm-Bawerk's Positive Theory of Capital, and in Chapter IV of William Smart's An Introduction to the Theory of Value. The tables are triangular because the most pressing want in one category typically is less pressing than the most pressing want in another category. An element in the space of commodities corresponds to the set of wants that the agent would choose to satisfy with the quantities of commodities specified by that element.

This mapping from quantities of commodities to sets of wants leads to a redefinition of marginal utility, which one might as well designate by a new name - marginal use. The marginal use of a quantity of commodity is, roughly, the different wants that would be added, with a set union, to the set of wants satisfied by the the given quantities of commodities with that additional quantity of the given commodity. McCulloch shows that a ranking of wants in different categories can arise such that a measure does not exist for the space of sets of wants. (A measure in this sense is a technical term in mathematics, typically taught in courses in analysis or advanced courses in the theory of probability.) He argues that the Austrian theory of the marginal use is thus ordinal. Surprisingly, his argument implies that the law of diminishing marginal utility does not require utility to be measured on a cardinal scale.

I haven't read Ian Steedman's work on consumption, but I think I'll mention it here.

5.0 Choices from Menus
Another generalization of the textbook treatment is to examine how a preference relation can be built out of a more fundamental structure. Imagine the agent is presented with a menu, where a menu is a nonempty set of elements of the commodity space. The agent is assumed to have a choice function, which maps each menu to the set of best choices, in some sense, in that menu. The agent is not postulated to rank either the elements not chosen for a given menu or the elements in the choice set.

A question: what constraints need to be put on choices out of menus such that preferences exist? Since a choice function can be constructed for which no preference function exists, some such constraints exist. I previously noted literature drawing on the logical structure of social choice theory in this context. Alan Isaac emphasizes temporal and menu independence in his overview of abstract choice theory.

6.0 Experimental Economics
I am emphasizing theory. A literature exists on experiments, many of which have falsified the textbook treatment of economics.

7.0 Computatibility, Conservation Laws, Etc.
Some of the above extensions of the textbook treatment seem to postulate some sort of structure within the agent's mind. Computers provide an arguable metaphor of mental processes, and some literature applies the theory of computability to economics. Gerald Kramer, for example, shows that no finite automaton can maximize utility in the simplest setting. I gather others have shown that the textbook treatment postulates that each agent's computation powers exceed those of a Turing machine, that agents compute functions that are, in fact, noncomputable. I turn to Kumaraswamy Velupillai's work for insights into computability, constructive mathematics, and economics. Philip Mirowski is always entertaining. One might also mention the literature on Herbert Simon's notion of satisficing

8.0 Conclusion
This post is a brief overview of some of what would be treated in a survey of variations and approaches to utility theory. Apparently, the notion of economic man can be complicated.

An Incomplete List of References
  • Colin F. Camerer (2007) "Neuroeconomics: Using Neuroscience to Make Economic Predictions", Economic Journal, V. 117 (March): C26-C42.
  • Alan G. Isaac (1998) "The Structure of Neoclassical Consumer Theory"
  • Daniel Kahneman and Amos Tversky (1979) "Prospect Theory: An Analysis of Decision under Risk" Econometrica, V. 47, N. 2 (March): pp. 263-292
  • Gerald H. Kramer () "An Impossibility Result Concerning the Theory of Decision-Making", Cowles Foundation Paper 274
  • Kevin J. Lancaster (1966) "A New Approach to Consumer Theory", Journal of Political Economy, V. 75: pp. 132-157.
  • J. Huston McCulloch (1977) "The Austrian Theory of the Marginal Use and of Ordinal Marginal Utility", Journal of Economics, V. 37, N. 3-4: pp. 249-280.
  • Judea Pearl (1988) Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, Morgan Kaufmann
  • Leonard J. Savage (1954, 1972) The Foundations of Statistics, Dover Publications
  • Chris Starmer (1999) "Experimental Economics: Hard Science or Wasteful Tinkering?" Economic Journal, V. 109 (February): pp. F5-F15
  • Ian Steedman (2001) Consumption Takes Time: Implications for Economic Theory, Routledge
  • S. Abu Turab Rizvi (2001) "Preference Formation and the Axioms of Choice", Review of Political Economy, V. 13, N. 12 (Nov.): pp. 141-159
  • John Von Neumann and Oskar Morgenstern (1953) Theory of Games and Economic Behavior, Third Edition, Princeton University Press


Anonymous said...

Interesting post by the way. Have you ever read Jonathan Barzalai's papers critiziing the foundations of utility theory on mathematical grounds? For example:

The claim that modern economic theory can be founded on indifference curves of ordinal utility functions is based on errors. For the same reasons that the mathematical theory of thermodynamics cannot be founded on ordinal temperature scales, modern economic theory cannot be founded on ordinal data. Economic theory should be corrected
using utility scales that enable the “powerful weapons” of algebra and calculus.

Robert Vienneau said...

I hadn't read that before. I guess my post is vulnerable to that critique. I see that whether or not a utility function is differentiable is meaningless if utility is only ordinal. The question, I guess, is whether indifference surfaces can be meaningfully defined for any given level of utility. And whether the differentiability of such indifference surfaces is meaningful.