You are on page 1of 18

AMARTYA SEN

INCOMPLETENESS AND REASONED CHOICE

The subtitle of Isaac Levis book, Hard Choices, explains the nature of the problem that he addresses in that classic work: Decision Making under Unresolved Conict.1 We have learned greatly from Levis analyses of why, despite our best efforts, the valuational conicts that we face may not always be fully resolved when the point of decision making comes, and how we may nevertheless use systematic reasoning to decide what one should sensibly do despite the presence of unsettled conicts. Indeed, through a variety of contributions stretching over several decades, Isaac Levi has powerfully illuminated the challenges of decision making in the presence of imperfect information, conicting evidence, divergent values, discordant commitments, and other sources of internal dissension.2 I seize the wonderful occasion of celebrating Isaac Levis work and accomplishments by presenting a series of observations on rational decision making with incompletely resolved internal dissensions. In that context, I comment also on the nature and use of incomplete valuational orderings and the ways and means of extending their reach. I pursue these issues in the form of addressing a series of questions. Sometimes I draw on Levis work, and at other times, I comment on differences that we may still have. As will be obvious, even when we disagree, my understanding of these issues is strongly inuenced by Levis thinking.

1. IS NONCOMMENSURABILITY THE PRIMARY REASON FOR


UNRESOLVED CONFLICT ?

Noncommensurability of different types of values is often seen as a reason indeed the main reason for incompleteness of an overall (all things considered) ranking. Indeed, the belief that commensurability is necessary for arriving at a completely ordered ranking of different options seems to have considerable following in the literature. Since I have always had some problem in grasping the reasoning behind that belief, I should perhaps try to articulate where my difculty lies.
Synthese 140: 4359, 2004. 2004 Kluwer Academic Publishers. Printed in the Netherlands.

44

AMARTYA SEN

Commensurability of two distinct objects stands for their being measurable in terms of each other. Noncommensurability is present when several dimensions of value are irreducible to one another.3 In the context of evaluating a choice, commensurability requires that in assessing its results, we can see the values of all the relevant results in exactly one dimension measuring the signicance of distinct outcomes in a common scale so that in deciding what would be best, we need not go beyond counting the overall value in one homogeneous metric. Since the results are all reduced to one dimension, we need do no more than checking which option will give us how much of the one good thing to which every value is reduced. We are, certainly, not likely to have much difculty in choosing between two alternative options both of which offers just the same good thing, but one offers more than the other. This is an agreeably trivial case, but the belief that whenever the choice problem is not so trivial, we must have great difculty in deciding what we should sensibly do seems peculiarly feeble (it is tempting to ask, how spoilt can one get!). Indeed, if counting one set of real numbers is all we could do, then there would not be many choices that we could sensibly and intelligently make. Whether we are deciding between buying different commodity baskets, or making choices about what to do on a holiday, or deciding whom to vote for in an election, we are inescapably involved in evaluating alternatives with noncommensurable results. Noncommensurability can hardly be a remarkable discovery in the world in which we live. And it need not, by itself, make it very hard to choose sensibly. For example, a ne mango may give us nutrition as well as some palatal or olfactory pleasure, whereas buying the record of a good song may offer a very different reward (not immediately reducible into the dimensions of the other), and given a budget constraint we could quite possibly face the choice of having one or the other. This involves choosing between noncommensurable results. And yet we may have no great difculty in opting for the mango when immensely hungry or starved, and going for the song, when well endowed with tasty food but short of melodious entertainment. The choice need not be hard to make in many situations, despite the noncommensurability involved. The distinct dimensions of values may not be reducible into one another, and yet there may be no problem whatsoever in deciding what one should sensibly do when our priorities or weights over these values are clear enough. Making choices with noncommensurable rewards is like speaking in prose. It is, in general, not particularly hard to speak in prose (even if M. Jourdain in Molieres Le Bourgeois Gentilhomme may marvel at our ability to perform so exacting a feat). But this does not negate the recognition

INCOMPLETENESS AND REASONED CHOICE

45

that speaking can sometimes be very difcult (for example, when one is overwhelmed by emotions), but that is not because expressing oneself in prose is in itself arduous. The presence of noncommensurable rewards only indicates that choice decisions will not be trivial (reducible just to counting what is more and what is less), but it does not at all indicate that it is impossible or even that it must be particularly difcult. What we have to ask is: why are some choices hard? This must be the case when the diverse values involved are both in conict and difcult to weigh relative to each other, that is, when (as Isaac Levi discusses with such care) it is hard to resolve their conicting pulls. The exercise of overall evaluation may not only involve more than counting, the process of non-trivial aggregation may be complex and challenging in some cases in a way it may not be in others. To say that noncommensurability is the primary cause of incompleteness amounts to missing the specic causes of incompleteness in favour of an ever-present precondition that has little discriminating relevance. To use an analogy, it would be like saying that we feel hungry primarily because we have a stomach. Certainly, it is hard to explain hunger without presuming something about the stomach, but it is not particularly useful to concentrate on the existence of the stomach in the human body in trying to explain hunger in the world. While the stomach is generically involved, we are likely to get more help in explaining hunger if we try to examine instead how difcult it may be for a person particularly a poor person to ll his or her belly. Similarly, even though the presence of noncommensurable results is generically involved in incompleteness, we have to investigate whether it is difcult or easy to weigh the different types of values involved and to resolve their conicts.
2. IS THERE ULTIMATELY ONLY ONE SOURCE OF IMPORTANCE ?

Noncommensurability does not in itself take us very far in understanding the nature of the decisional problem to be addressed. It is, therefore, in many ways a pity that so much attention has been heaped in the literature on the existence of noncommensurability as such, which is an omnipresent and ordinary feature of the world, compared with the distinct reasons that make value conicts so serious in some cases and not at all in others. But perhaps there is a different, more dialectical, explanation of why several perspicacious thinkers have chosen to emphasize the importance of noncommensurability. The attention paid to noncommensurability is partly a critique of the belief often implicit that there must ultimately be only one source of signicance. By focusing on noncommensurability, the critics of such reductionism assert something of importance regarding the

46

AMARTYA SEN

absence of a single standard of value.4 Indeed, the hold of the tradition of reducing everything to one homogeneous virtue may explain why the mundane recognition that there are different values that are not reducible to each other has been seen as worth asserting. Even a humdrum cognizance may have a positive and indeed a constructive role in the dialectics of conceptual diagnosis. Adam Smith complained more than two hundred years ago about the tendency of some philosophers (he had separated out Epicurus, but clearly had others in mind as well, including circumstantial evidence indicates his friend, David Hume) to look for a single homogenous virtue in terms of which all values could be explained:
By running up all the different virtues to this one species of propriety, Epicurus indulged a propensity, which is natural to all men, but which philosophers in particular are apt to cultivate with a peculiar fondness, as the great means of displaying their ingenuity, the propensity to account for all appearances from as few principles as possible. And he, no doubt, indulged this propensity still further, when he referred all the primary objects of natural desire and aversion to the pleasures and pains of the body.5

There are indeed schools of thought which insist, explicitly or by implication, that all the appearances of value must be reduced ultimately to a single source of importance. This claim about the nature of value is often supplemented by the further thought, in which commensurability does come in, that in order to be able to choose rationally, all values must be reduced to one. Evidently, the protagonists of this view believe that human beings can count but cannot evaluate. The counting freaks (if I may call them that, without intending any disrespect) include some but not all utilitarians. There is indeed a version of utilitarian reasoning that takes the form of arguing that there is no noncommensurability at all in what ultimately matters, to wit utility. In the example considered earlier, if we contingently have no difculty in choosing between the mango and the song, it is because (the argument runs) both can be judged by their respective ability to generate utility. In the last analysis (the argument insists), we count, not judge. I shall not scrutinize here this particular argumental loop, but only note that utility may be far from a homogenous magnitude (a recognition that had not escaped that great utilitarian, John Stuart Mill), and also that there can be other reasons for choice (other than the pursuit of utility), no matter how utility is substantively dened. If, however, utility is not dened independently, but only as the real-valued representation of the binary relation underlying choice behaviour (as is common in modern economics), then utility is not only not one good thing, it is not a thing at all, but a mere phantom of representation. Furthermore, even a consistent phantom

INCOMPLETENESS AND REASONED CHOICE

47

may not exist when choice behaviour is non-binary. I have discussed these issues elsewhere, and will desist from pursuing them further here.6
3. HOW USEFUL IS LEVI S GENERAL CONCEPT OF V- ADMISSIBILITY IN UNDERSTANDING THE DEMANDS OF
RATIONAL DECISION MAKING WITH UNRESOLVED CONFLICTS ?

V-admissibility, which is a central concept in Levis investigation, is indeed very useful. Even though I shall presently argue (in Section 5) that particular questions can be raised about some applications of the idea of V-admissibility that Levi endorses, let me rst discuss why the concept is, I think, important and helpful. In a set of options (let me call it the menu), the V-admissible options are those that have not been prohibited by the agents value commitments from being chosen by the agent.7 They are, thus, admissible relative to the agents valuations of the feasible options as better or worse, all things considered. The concept of V-admissibility introduces, in two distinct ways, a useful gradualism in the process of systematic choice. First, if an option is not V-admissible, then it would not be sensible to choose it, but if it is V-admissible, there may still be further questions to be asked as to whether it is optimal or not. In being denitive in exclusion but only permissive in terms of inclusion, it makes the choice process capture an important asymmetry. Weeding out the clearly unchoosable options can be a good way to begin, but even when the rotten ones have been weeded out, the remainder may call for further and closer attention. Second, V-admissibility itself is a parametric concept: the criteria of admissibility are not already ingrained in the very idea of V-admissibility and they have to be additionally dened, and can be varied. As the criteria are more and more specied, the constraints imposed by V-admissibility can be gradually tightened. Sequential tightening can be helpful in coming to grips with hard choices. Let me illustrate by recharactering Levis investigation in the form of four distinct steps associated with the idea of V-admissibility (drawing on different parts of Levis work). (1) Single-dimensional valuation. It is, as discussed already, trivial to decide what to do when the different results are all fully commensurable. We can, in this very special case, simply count our way to checking which option or options are the best (that is, have most of the one good thing to which everything else is reduced). With single-dimensional valuation, V-admissibility is both trivial and decisionally denitive: in one step we can reject all the rejectable alternatives and go straight to what we

48

AMARTYA SEN

should choose.8 If this does not work and mostly it will not then we must get into non-trivial exercises. (2) Congruent multi-dimensionality. When distinct and disparate values all move together, in the same direction with each other, the ranking of options can simply follow the shared ordering of all the distinct value commitments. As Levi notes, when the agents value commitments generate no conict but constrain the agent to evaluate his feasible options in an unequivocal manner, the agent is obligated by those commitments to restrict his choice to one of the feasible options which is optimal according to the mandated ranking.9 The V-admissible alternatives are those that could not be eliminated on the ground of their being inferior to some feasible option in terms of each of the value commitments involved. This is also a simple enough case, but it does not presuppose any commensurability at all. (3) The Weighted Average Principle: So far nothing more than ordinal ranking of each value commitment has been invoked, but we must now go beyond that. Let v1 , . . . , vn be the n numerical valuational functions that give the value of each option according to each of the n value commitments respectively. It is assumed that each of these vi is cardinally measurable unique up to a positive afne transformation.10 Levi connes attention now to the weighted sum of all the vi as reecting the aggregate value v of the respective options, for a set of non-negative weights (w1 , . . . , wn ).11 So we have: v = wi vi , and depending on the weights we choose for the respective value commitments, we get a corresponding set of aggregate values of each option. If we had a uniquely specied set of such weights (w1 , . . . , wn ), then of course the valuational exercise would be over. But so long as the conict is unresolved, this could not be presumed. So, at this stage, Levi concentrates attention on the entire class of non-negative weights. However, some options may end up having a lower aggregate value than another under every permissible set of non-negative weights, and in that case, the uniformly lower valued alternative can be eliminated as being non-admissible. This way the afforced criterion of V-admissibility eliminates some non-admissible options in the menu.12 This is a critically important and innovative step in Levis reasoning, and I shall examine the pros and cons of going this way in Section 5, when I specically scrutinize the weighted average principle. (4) Categorical preference and optimality. It is possible to go beyond valuational conicts when aggregate valuations are in conformity with each other, at least in part. For example, if x is strictly preferred in value to y according to all permissible valuation rankings, then x is categorically preferred to y.13 The notion of categorical preference is useful in checking admissibility, but it can also in some cases with luck help to identity an

INCOMPLETENESS AND REASONED CHOICE

49

optimal choice. When it turns out that there exists an option x in the menu such that for every permissible valuational ranking, x is at least as good as every other alternative in the menu, then x is declared to be V-optimal. If, furthermore, the rankings are such that any option y in the menu that is regarded as indifferent to an optimal option x is also optimal (but y is denitely not optimal when x is ranked above y), then we get to the more regular idea of full optimality rather than only V-optimality.14 The possible existence of optimal or V-optimal options can make decisions easier to take despite unresolved conicts, but the broader idea of V-admissibility gets us, in general, part of the way, even when we do not arrive at a categorically justied optimal decision. In the gradualist approach investigated by Isaac Levi, the role of each of the nely dened concepts is explored and explained.15

4. IS THERE AN ALTERNATIVE SYSTEMATIC APPROACH TO RATIONAL


DECISION MAKING UNDER UNRESOLVED CONFLICTS ?

There is, and indeed, to a great extent, Levis strategy can be seen as a reasoned departure from an older approach that focuses on maximization based on a possibly incomplete ordering derived from the intersection of different value commitments. That approach can be seen in terms of its two constitutive components, viz. maximization and intersection.16 I shall call it (not terribly imaginatively) intersection maximization.17 The idea of maximization as choosing an option that is no worse than any other feasible option can work with incomplete orderings. To qualify as maximal an option need not be shown to be at least as good as all the other feasible options only that it is not strict worse than any. The idea of maximality in this broad sense has been well formalized by Bourbaki, Debreu and others.18 If decision making is based on maximization (rather than optimization choosing an option that is at least as good as every other option), it is not a requirement that all value conicts be resolved before a reasoned choice can occur. The incomplete ordering to be used for maximization can be derived on the basis of the intersection of the different valuational rankings. Even when unresolved conicts exist, an incomplete ordering can be identied that ranks two options in the overall ranking in a certain way if and only if those options are ranked in the same way by all the different values involved. So, if x is ranked above y according to each of these values, then x is ranked overall above y. Similarly, if x and y tie in terms of each of these values, then x and y are taken to be indifferent overall. The intersection partial ordering generated in this way can then be used, along with the approach of maximization, for systematic choice

50

AMARTYA SEN

despite the presence of unresolved conicts (which can make the overall ranking incomplete). The combined use of maximization and intersection is basically isomorphic with Levis identication of a mandated ranking related to V-admissibility before we get to his weighted average principle.19 Since the older approach does not invoke the weighted average principle, we can perhaps ask what alternative principle does it use? The answer is nothing: it stops right there with maximization based on the intersection partial ordering. In this sense, intersection maximization is an ineloquent theory (I shall presently comment on the strategic use of this lack of eloquence). The maximal set is the set of all options that are undominated by the intersection partial ordering. It can be shown that maximization, in this broader sense (broader than optimization, and also as it happens, broader than V-admissibility with Levis weighted average principle) yields many important and rather far-reaching properties in the discipline of choice, and it can be defended both on analytical and substantive grounds.20 The general approach of maximization can also incorporate various additional features that were not part of the old structure of maximality identied by Bourbaki and Debreu: such as incorporating actions of agents as a part of the comprehensive outcome of choice, which allows us to re- examine the dividing line between consequentialist and deontological reasoning. Various other properties can also be optionally imposed by utilizing the capacious format of maximization, and making use of the room for further articulation left open by the ineloquent form of that approach.21 We have to examine whether Levis addition of weighted average principle and its implications for categorical preference can also be sensibly added on to the basic structure of intersection maximization. I take up that issue next.
5. IS THE CRITERION OF WEIGHTED AVERAGE PRINCIPLE FOR
V- ADMISSIBILITY PERSUASIVE AND ACCEPTABLE ?

There are certainly good arguments in favour of using some additional structure, like Levis weighted average principle, to extend the reach of intersection maximization. There is, rst, a manifest need here, and second, Levis particular constructive proposal has some conspicuous merit. I begin with the issue of need and the motivation for Levis departure. The maximal set based on intersection can be quite large. Consider Levis example of a hapless ofce manager, called Jones: I shall call her Ms. Jones. She wants to hire a secretary who is a good typist and a good stenographer as well (and thus has two value commitments), and con-

INCOMPLETENESS AND REASONED CHOICE

51

siders three candidates: Jane, Dolly and Lilly. Jones nds that in terms of typing skill, Jane is better than Dolly and Dolly superior to Lilly, but in stenographic competence, Lilly is better than Dolly, who is better than Jane. The intersection of these two rankings is empty, and intersection maximization based on these two orderings would suggest that all three are maximal. Levi, not surprisingly, sees this as hopelessly inarticulate. He wants to ask such questions as whether Dolly, the middle skilled in each, is the second best typist and second best stenographer, or merely the second worst typist and second worst stenographer. Dolly is, of course, all these, in ordinal terms, but Levi wants to be able to get and use more information about where Dolly gures vis-a- vis Jane and Lilly. This takes Levi to cardinal valuational functions vi and then to the weighted average principle. With this motivation for wanting a more informed and through that a more articulate choice, I am in total sympathy.22 The issue to be examined is whether the particular procedure developed by Levi is adequate. The weighted average principle is a way of cutting down the intersection-based maximal set by eliminating some options that would have a lower value than an alternative option in terms of all non-negative weights that can be placed on the cardinal valuational functions associated with the respective value commitments. The Levi approach goes beyond the older maximization approach both (i) by introducing cardinal measurement associated with each value commitment, and (ii) by using the strategy of weeding out any option as V-inadmissible if it gets dominated by all possible weighted aggregate values by some other feasible option (in terms of the cardinal representation chosen). This is where the departure comes, and we have to ask whether it is convincing. The second argument, which takes us beyond motivation, is that Levis reasoning certainly has considerable plausibility (which as I shall presently discuss is not the same thing as its being, all things considered, convincing). The plausibility can be illustrated by giving valuational numbers to the two skills of the three candidates each. Consider the following valuations:
Candidates Jane Dolly Lilly Typing skill 10 2 1 Stenographic skill 1 2 10

Given the cardinal characteristics of this measurement, any positive afne transformation of these numbers will serve as well.

52

AMARTYA SEN

We can add up the two numbers of each with any set of non- negative weights. But it is clear that no matter what these chosen weights are, Dolly will get trumped by either Jane or Lilly, since Dolly is nearly as bad a typist as Lilly and nearly as bad a stenographer as Jane. To see this consider the possibility that Dolly gets a higher total score than Jane. In that case the weights must be such that the one- point advantage in stenography that Dolly has over Jane (with her mark of 2 against Janes 1) outweighs the 8-point advantage that Jane has over Dolly in typing (with their respective gures of 10 and 2). So the weight on stenographic skill must be at least 8 times as great as on typing skill. But, then, clearly Lilly will beat Dolly hollow with her 8-point advantage over Dolly in stenography, despite her one-point decit vis-a-vis Dolly in typing. If Dolly beats Jane, then Lilly beats Dolly, and it can also be readily checked that if Dolly beats Lilly, then Jane will beat Dolly. Dolly is, thus, not V-admissible, given Levis weighed average principle. The reasoning not only takes us beyond the purely ordinal comparisons of intersection maximization, it also has some evident appeal. Though I have not been retained by luckless Dolly, let me now argue, rst, against this particular conclusion, and then, against the principle of weighted average in general. The central issue is this: can we do an overall evaluation of the candidates without asking how the importance of the two skills may vary as we consider different levels of achievement of the candidates? Suppose that a skill level of 1 in typing, which Lilly has, does not make it possible for letters to be typed well enough to be despatched, but level 2, which Dolly has, makes that possible, though it is nowhere near the superb level of typing skill that Janes mark of 10 indicates. Suppose also that a skill level of 1 in stenography, which Jane has, leads to total chaos in the ofce (say that a third time please, Ms. Jones), whereas Dollys level 2, modest as it is, prevents that, even though the stenographic work would be massively better with a skill level of 10, which Lilly enjoys. Faced with these considerations, Ms. Jones may drop Jane to prevent chaos in ofce, and she may also let Lilly go so that some letters can actually get typed. Dolly satises the qualifying level in each skill, and may well be chosen on these grounds (despite her V-inadmissibility in terms of the weighted average principle). I rest my case for Dolly or more accurately for not giving Dolly the boot without checking what the importance of the different levels of skill are for the choice at hand (and not just how much skill there is). I turn now to the more general issue of the acceptability of the weighted average principle. Underlying my scepticism is a question about the inter-

INCOMPLETENESS AND REASONED CHOICE

53

pretation of the valuational numbers related to each value commitment. What do the numbers given by vi stand for? Do they: (1) measure the levels of accomplishment in eld i, or (2) measure the importance of these accomplishments for the purpose for which the choice decision is being taken? Levis language suggests that he takes the former interpretation, so that the numbers represent the levels of skill of Dolly and others. If so, we still have to ask how valuable is that skill for the decision at hand, and this cannot be answered independently of the levels of that skill and the presence of other factors that allow or facilitate or hinder the use of that skill. For example, Janes excellent typing skill (mark 10 no less) can be quite valueless in terms of Ms. Joness decisional problem if the woeful nature of her stenographic skill makes her an impossible holder of the ofce position in question. We need not only a valuation of the respective skills (and more generally, of the extent of accomplishment in each value commitment), but also an evaluation of the contingent importance of the exact level of skill (or accomplishment) given everything else that is involved in the decisional picture. If, on the other hand, the second interpretation is taken, then we have to ask how can the importance of a skill (or more generally, of a value accomplishment) for the decision at hand be represented independently of other factors in the choice? How can we, for example, determine that the choice- context importance of Janes typing skill, in an additive framework, is invariably 10 for this ofce job, no matter whether she can do anything else at all that may also be a part of the job in question? Something or other is clearly missing in this framework, and the operation of weighted addition, which leads the way to the weighted average principle, can be deeply problematic. Perhaps an analogy from social choice theory can help to clarify the issue. Consider a welfarist framework, in which the social value of a state of affairs is seen as a function of the welfare levels of all the people in that state.23 We can think of each persons well-being as a kind of value commitment, in terms of an analogy with Levis framework. To do this exercise, we need: (1) to evaluate each persons well-being (measured on its own); (2) to make interpersonal comparisons of the values of different persons well-being (putting them in a comparable scale); (3) to decide on the weights to be put on increments or diminutions in each persons well-being respectively vis-a-vis those of others (reecting

54

AMARTYA SEN

what importance we want to attach respectively to distributive and aggregative considerations).24 These distinct exercises cannot be combined in one assignment of vi functions (unique up to a positive afne transformation) which are then simply added together with non-negative weights. No matter how we choose the values of vi functions, something or other would not have been addressed. I would argue that a similar lacuna exists in the framework of V-admissibility with vi functions, limited to the class of xed afne transformations and xed non-negative weights. Judging the decisional importance of the accomplishments of different value commitments requires more than a weighted averaging of the different accomplishments.

6. CAN WE USE V- ADMISSIBILITY WITHOUT THE WEIGHTED


AVERAGE PRINCIPLE ?

Even though the particular proposal of using the weighted average principle with V-admissibility seems problematic that does not negate the motivation that led Isaac Levi to seek an extension of the reasoning underlying intersection maximization. The maximal set given by the intersection of all the value commitments can be unhelpfully large, and it is sensible enough to try to see how that set can be reduced in size to give more bite to the decisional process. Dolly may be hard to eliminate from the list of three through the weighted average principle in particular, but that need not indicate that all three must be seen as ne appointees merely because the intersection partial ordering is empty. Levis motivation in seeking more stringent criteria in the form of tighter requirements of V-admissibility is in general just right. Consider a more general specication of the problem of going beyond intersection maximization through V-admissibility, by combining Levis motivation with formal investigations pursued in social choice theory.25 We follow Levi in considering valuation functions vi related to each value commitment, but do not necessarily constrain the uniqueness properties of each vi to the class of positive afne transformations. Depending on the type of values involves, the class may be wider (and the extent of measurability correspondingly less, e.g., ordinal), or narrower (and correspondingly have more measurability, e.g., ratio scale), and it is possible also to consider various intermediate possibilities of a hybrid kind.26 We can proceed from there to consider aggregation functions v = fk (v1 , . . . , vn ), and impose invariance conditions that reect the particular extents of measurability and any comparability restrictions that we may want to

INCOMPLETENESS AND REASONED CHOICE

55

impose.27 We can consider a number of such aggregation functions fk , with k = 1, . . . , m, in a class F. If for some feasible y, for all fk in F, v(x) is less than v(y), then x is not V-admissible. This general structure can accommodate Levis weighted average principle if the valuation functions vi are cardinal and if the aggregation functions fk in F are restricted exactly to the class of weighted averages. Similarly, by choosing ordinal vi functions and an ordinal class F of aggregation functions fk , we can get back to simple intersection maximization. But there are a great many other possibilities as well, and V-admissibility can be used to consolidate the minimal articulation of intersection maximization, and then to go beyond that to the extent that is permitted by our actual ability to marshall information and to scrutinize our assessments. Unresolved value conicts leaves open many possibilities of methodical use of decisional reasoning, and V-admissibility can serve as a general format that accommodates as much articulation as we can reasonably justify.
7. IS INCOMPLETENESS ALWAYS TENTATIVE , AWAITING
RESOLUTION ?

In the preceding discussion the focus has been on reducing incompleteness as much as possible. But what about any remaining incompleteness? Should we see it as an embarrassment, or as a defect that calls for ways and means of total elimination. I would argue that the answer must depend on the nature of the remaining incompleteness. In many cases the incompleteness is best described as tentative. It awaits resolution (with more information, or deeper analysis, or closer scrutiny, or whatever), whether or not the resolution actually occurs. This kind of incompleteness has to be contrasted with the idea of assertive incompleteness.28 This category separates out cases of incompleteness in which the lack of completeness is positively asserted, yielding statements such as x and y cannot be ranked. It is radically different from incompleteness that is tentatively accepted, while awaiting or working for completion. The partial ranking may simply not be completable, and may not even be ideally completed. Rather, incompleteness may be the right answer in these cases. It is useful to distinguish between three different types of assertive incompleteness related to the source of the incompletability involved. First, when the inquiry concerns a specic eld of ethical or decisional judgment (such as justice), the assertive incompleteness may relate to the domain of that type of judgment (for example, the reach of a theory

56

AMARTYA SEN

of justice). The recognition that incompleteness of judgments can be a constituent part of the denite conclusions advanced by a complete theory of justice can be quite momentous for practical reason. Indeed, even a complete theory of justice can yield and assert incomplete rankings of justice. Afrmed incompleteness (e.g., x and y cannot be ranked in terms of justice) may be as much a denite conclusion as the decision that x and y can be ranked in terms of justice (and in particular that, say, x is more just than y, or that x is just, whereas y is not). However, it must also be noted that even when it is assertively concluded that x and y cannot be ranked in terms of justice, this does not entail that they cannot be ranked in terms of some other type of ethical or political concern. There is a distinction here between this kind of eldspecic conclusion (important as it may be) and more ambitious claims of assertive incompleteness for all things considered decisions. Second, even for an all things considered evaluation, a partial ranking (or a partial partition) may be a crucial assertion as the end product at a particular stage of a multi-stage exercise. A theory may assert incompleteness for a well-dened purpose, leaving room for a possible extension through an appropriate subsequent stage. If some decisional issues are not decidable by, say, general (or foundational) ethical reasoning, the incomplete ranking emerging from general ethical reasoning may be an appropriate assertion for that general stage. This may perhaps be supplemented by a subsequent stage using other procedures (for example, some kind of a democratic decision mechanism) that extend the general ethical reasoning by choosing between alternative courses of action that are all consistent with general ethical reasoning.29 The afrmation of incompleteness at one stage of the process may be critically important for the necessity and viability of the next stage. Third, aside from eld-specic and stage-specic assertive incompleteness, there is the possibility of an unqualied assertion of incompleteness. It is possible that incompleteness may be a durable and denitive part of the end product of all things considered and all stage included evaluation. It may be as far as we can proceed with reasoned discrimination, given the information that we can conceivably have. If so, the incompleteness will not await completion at a later stage or over a wider eld of reference, and will yield such statements as: x and y denitely cannot be ranked for decisional purposes.30 There is a need to see assertive incompleteness as a conceptual category of its own. For example, a ne-grained discrimination between two value commitments with exactly specied weights may simply be beyond the reach of reasoned ethical scrutiny, or may demand information that may not only be contingently

INCOMPLETENESS AND REASONED CHOICE

57

lacking but also impossible to obtain even in principle. For example, it can be argued that interpersonal comparisons of well-being, though by no means impossible, may not take us all the way to invariance conditions that yield one-to-one correspondence of everyones well-being numbers vis-a-vis each other.31 I end by noting that the recognition of the possibility of assertive incompleteness does not reduce in any way the value of scrutiny and investigation aimed at reducing the extent of tentative incompleteness through continued scrutiny of unresolved conicts. Nor, for that matter, does it rule out the possible usefulness of challenging whether a putative claim of assertive incompleteness is indeed justied. Assertive incompleteness no less than tentative incompleteness is well within the domain of inquiry and justication which Isaac Levi has done so much to clarify for us. We know from Isaacs work how magnicently capacious that domain is.

NOTES
1 Isaac Levi, Hard Choice: Decision Making under Unresolved Conict, Cambridge

University Press, Cambridge (1986).


2 Isaac Levi, Gambling with Truth, Knopf, New York (1967); Information and Inference,

Synthese 17, 369379 (1967); On Indeterminate Probabilities, Journal of Philosophy 71, 391 418 (1974); The Enterprise of Knowledge, MIT Press, Cambridge, MA (1980); Hard Choices (1986); Conicts and Inquiry, Ethics 102, 814834 (1992); The Covenant of Reason, Cambridge University Press, Cambridge (1997). Levi has been concerned with decisions in epistemology as well as practical reason. The focus of this essay is, however, exclusively on the latter, even though some of the basic issues have relevance to epistemology as well. 3 The Covenant of Reason (1997, p. 236). 4 As Isaac Levi explains, when the perspective on inquiry and justication is combined with a rejection of a single xed standard of value, the key elements of the pragmatism of Peirce and Dewey which I admire are identied (The Covenant of Reason, p. 218). 5 Smith, The Theory of Moral Sentiments, revised edn., VII.ii.2.14 (1790). Republished, Clarendon Press, Oxford (1976, p. 299). 6 See Choice, Welfare and Measurement, Blackwell, Oxford (1982); Harvard University Press, Cambridge, MA (1997); On Ethics and Economics Blackwell, Oxford (1987); and Rationality and Freedom, Harvard University Press, Cambridge, MA (2002). 7 Levi, Hard Choices (1986, p. 10). 8 I am abstracting here from the problem of innite sets in which even a complete ordering need not necessarily yield a best or an optimal alternative. That raises a different range of issues, with which I shall not be concerned in this essay. In the present investigation, we may as well assume that the set of options is nite, so that a complete ordering will always identify a best alternative. Indeed, even an acyclic but complete ranking will do that; on this and related results, see my Collective Choice and Social Welfare, HoldenDay, San Francisco (1970). Republished, North-Holland, Amsterdam (1979, Chap. 1 ),

58

AMARTYA SEN

and Choice Functions and Revealed Preference, Review of Economic Studies 38 (1971); Hans Herzberger, Ordinal Preference and Rational Choice, Econometrica 41 (1973); and Levi, Hard Choices (1986). 9 Levi, Hard Choices (1986, p. 9). 10 An afne transformation is of the form: a + b.v , and the positivity of such a transi formation refers to b being positive. On the characteristics and uses of aggregations based on cardinal values (with uniqueness up to positive afne transformations), see my Collective Choice and Social Welfare (1970); Kotaro Suzumura, Rational Choice, Collective Decisions and Social Welfare, Cambridge University Press, Cambridge (1983); Levi, Hard Choices (1986). 11 The weights w , . . . , w are not only non-negative each, but also they must sum to 1. n 1 12 Levi, Hard Choices (1986, Section 5.4, pp. 7779). See also the motivating discussion on pp. 6977. 13 This may require us to go beyond the weighted average principle. An extreme case would be one in which there is a unique set of appropriate weights (w1 , . . . , wn ). But categorical preferences over a subset of alternatives can arise in many different ways, and may even be precipitated, in some cases, by the weighted average principle. 14 Levi, Hard Choices (1986, pp. 8355). Levi discusses these structures in the context of examining values revealed by choices, but they throw light also on the movement from values to choices, and not merely from choices to revealed values. 15 There are many other structural features which Levi has investigated, which I shall not examine in this essay. 16 Combined use of maximization and intersection can be found in many exercises in public economics. See for example A. B. Atkinson, The Measurement of Inequality, Journal of Economic Theory 2 (1970); Amartya Sen, On Economic Inequality, Clarendon Press, Oxford (1973); extended edition, with a joint Annexe with James Foster (1997). 17 In On Economic Inequality (1973) it was called the intersection approach, taking maximization for granted. 18 N. Bourbaki, Elements de Mathmatique, Herman, Paris, and Theory of Sets, AddisonWesley, Reading, MA (1968); Gerard Debreu, The Theory of Value, John Wiley, New York (1959). 19 Levi, Hard Choices (1986, p. 9). 20 On this see my Maximization and the Act of Choice, Econometrica 65 (1997). 21 Some of these possibilities have been explored in my Maximization and the Act of Choice (1997) and Consequential Evaluation and Practical Reason, Journal of Philosophy 97 (2000). 22 I have a natural sympathy here for various reasons, including the fact that I had a similar motivation in trying to move social choice theory towards accommodating more information on cardinality and interpersonal comparability of well-being and also including more cognizance of freedoms and liberties, in Collective Choice and Social Welfare (1970). 23 This is, of course, a very limited model of social choice, but adding further considerations, such as freedoms, liberties or rights, will not make the problem at hand any easier. 24 For the distinctions involved see my Collective Choice and Social Welfare (1970) and Choice, Welfare and Measurement (1982). 25 For assessments of alternative contributions to the ways and means of using more information in social evaluation, see my Social Choice Theory, in Kenneth Arrow and

INCOMPLETENESS AND REASONED CHOICE

59

Michael Intriligator (eds.), The Handbook of Mathematical Economics, North-Holland, Amsterdam (1986); and Kenneth Arrow, Amartya Sen and Kotaro Suzumura (eds.), Social Choice Re-examined, Elsevier, Amsterdam (1997). 26 See my Collective Choice and Social Welfare (1970, Chaps. 7 and 8 ). 27 The form of fk may or may not be additive. 28 I have discussed the notion of assertive incompleteness in Maximization and the Act of Choice, Econometrica 65 (1997); and in Justice and Assertive Incompleteness, Rosenthal Lectures (Lecture 2), Northwestern University Law School (1998). 29 The fact that despite my arguing for the use of the capability perspective in comparing individual advantages and my attempt to highlight the relevance of some basic capabilities in particular (for example in Well-being, Agency and Freedom: The Dewey Lectures 1984, Journal of Philosophy 82 (1985)), I have failed to specify a xed list of distinct capabilities (with specied weights or other ways of prioritization) has been the source of some chastisement I have received. However, if such a xed list with xed priorities and xed weights were indeed arrived at by general ethical reasoning, it is not clear to me how this would be consistent with the democratic process of setting priorities and precedence. 30 Assertive incompleteness must not be confused with an assertion of indifference. If it is claimed that x and y cannot be ranked, then that is what it says, not that they can be ranked as equals. Indeed, in some ways incompleteness is the opposite of indifference. Consider two possible claims: (1) x is at least as good as y, and (2) y is at least as good as x. If x and y are indifferent, then both (1) and (2) are true, whereas if their ranking is assertively incomplete, then (1) and (2) are both denied. On this distinction, see my Collective Choice and Social Welfare (1970, Chap. 1 ), and Maximization and the Act of Choice (1997). 31 I have tried to discuss the class of limited though possibly quite extensive informational discrimination in Interpersonal Aggregation and Partial Comparability, Econometrica 38 (1970), and in Collective Choice and Social Welfare (1970), and also in On Weights and Measures: Informational Constraints in Social Welfare Analysis, Econometrica 45 (1977). Trinity College Cambridge, CB2 1 TQ U.K. and Harvard University Cambridge, MA 02138 U.S.A.

You might also like