28 ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT On Intertemporal Preferences in Continuous Time The Case of Certainty by Chi-fu Huang and David Kreps WP //2037-88 July 1988 MASSACHUSETTS INSTITUTE- OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE. MASSACHUSETTS 02139 On Intertemporal Preferences in Continuous Time The Case of Certainty by Chi-fu Huang and David Kreps WP //2037-88 July 1988 On Intertemporal Preferences in Continuous Time: The Case Chi-fu of Certainty * Huang! and David Krepsf Revised, June 1988 on the space of consumption patterns in a continuous time setting are discussed. A family of topologies which give an economically reasonable sense of closeness and have an appropriate intertemporal flavor is suggested. Strictly concave time-additive utihty Different topologies certain functions are not continuous in this topology, since they yield preferences with the property that consumption at nearly adjacent dates are perfect The topological duals of our suggested topologies are spaces of absolutely continuous functions whose derivatives satisfy certain "boundedness" restrictions. A utility function that is continuous in the suggested topologies and that captures an intuitively appealing notion of time complementarity of consumption is provided. non-substitutes. 1. INTRODUCTION AND SUMMARY Consider an economic agent pose there is zero and one. a single We who lives from time ask the following questions: First, to time is at if we = t 1 . Sup- any time between how might we his lifetime? Second, space of possible consumption patterns, what is, = consumption commodity consumable agenf s consumption pattern over that < let represent the X^ denote the an appropriate topology on X^ ; one that gives an economically reasonable sense of "closeness", possess- ing the appropriate intertemporal flavor and that, at the same time, is relatively well-behaved mathematically? Third, what form will equilibrium prices take if they are given by a continuous and increasing linear functional? Economic and questions. financial theorists Consumption patterns have long "known" the answers are represented to these by nonnegative real-valued The authors would like to thank Peter Diamond, Darrell Duffie, Oliver Hart, Larry Jones, Andreu Mas-Colell, Scott Richard, and two anonymous referees for comments. Itzhak Katznelson told us about Orlicz Spaces. Tong-sheng Sun • pointed out an error in an earlier version. f School of Management, Massachusetts Institute of Technology J Graduate School of Business, Stanford University 1 . . functions of time, c{t), where c{t) represents the rate of consumption at time t Two . such consumption patterns functior\s; say, in the sup-norm topxjiogy or represented by maximization of U{c) Arrow and Kurz of come from L« We , and c(t) = c{i) are "close" if in V some J^ u{c{t),t)dt , they are close as topology. Preferences are where, in the terminology (1970), u is the felicity function of the individual. And prices for the appropriate value of q are dissatisfied with these standard answers. The causes of our we dis- satisfaction will be recorded in greater detail below, but basically consumption one time should be something of a substitute for consumption at That near-by times. at other, an agent for is, a while, until, say, a large meal time .5001 (not at finite rate) should be possible. . The at time, say, .5, feel that should sustain Consumption taken in large gulps gained from frequent equally utility spaced and equal sized gulps should approximate that gained from continuous ingestion at the corresponding average rate, as should consumption at quickly varying rates. And, in consequence, prices for consumption at dose-by times should be approximately the same. of the previous paragraph that will We set out to find answers to the questions conform to these intuitions, while at the same time retaining mathematical properties important for economic analysis. We begin in Section 2 with a formulation of the consumption set and con- sumption spaces. For the set of feasible consumption patterns, of nonnegative, nondecreasing, right continuous with on [0,1], denoted X^. For given x G A'+ and cumulated consumption from time zero to time (That is, in terms of standard treatments, x{t) model both continuous ingestion and gulping. the space of consumption bundles the space of functions of The space A" bounded is the linear variation comes equipped with a on t e left limits we use the space (RCLL) functions [0,1], x{t) denotes the ac- under consumption pattern x t = c{s)ds) J^ . speak of net trades, In order to span of [0, 1] number X+ This permits us to . Denoted A' , this is that are right continuous. of topologies, any of which could conceivably serve for our purposes. In order to investigate their we construct in Section 3 a "wish Our wish keeping list in comes mind in two we end, that, in the of properties for a suitable topology. list" parts. First we paper in terms of economics; when different Once we have our wish This fare. done is tially. is We is and Jones ||x|| we = will use J^ \x{t)\dt denoted hereafter by topologies are all weak-+ topology with fail rather substan- respect to continuous we [1984].) endowed with we But, as desire, on economies where the commodity this topology. will see, the find appealing topolgies have all 5, to a family of norm one representative member of + T \x{l)\ .) We for x will e X show ; (See weak topology Mas- lacks a topologies. (For this inthis family, the topology induced in section 5 that, topologically equivalent to the different uniformities, nomic well the standard topologies quickly look at the topologies induced by the a substantial literature turn instead, in Section troduction, we how uniform property that we find appealing. certain is the the heart of this be (uniformly) "close" together. Skorohod topology; these two (i.e., is are giving our intuition concerning can see the space of positive measures, Colell [1975] norm We to the story, and more importantly, analysis. Then, we will we add uncertainty under the sup norm) has many of the properties [0, 1] and, indeed, there space the The weak topology functions on list, in Section 4. norm and total variation essentially, consumption patterns to The discussion here desirable economic properties. list give useful mathematical properties, want will do optimization, and conduct equilibrium we suitability, weak on A'^. given by the this norm these norm by , topology. But they induce and our norm topologies give the uniform property which and which the weak topology the properties we find desirable fails to provide. These norm from the point of view of eco- intuition and, in addition, they are mathematically very well behaved; for example preferences continuous representations. in these topologies admit (continuous) numerical we In Section 6, nomic models (the integral of felicity of T are continuous in of consumption felicity of ask whether the usual representation of preferences in eco- is, at In Section 7, ^ if and only if rates) gives preferences that the marginal felicity of the rate each point in time, constant. (That is, rate function for time t, u{c{t),t) = consumption functions a and does so It . consumption u{c{t),t) if a{t)c{t) + is the I3{t) for .) we turn to characterization of the topological dual spaces for our topologies. In particular, the dual (X,T) (denoted, to the space of Lipschitz continuous functions. In this section, hereafter, as we X*)\s also confirm some desirable mathematical properties of the dual spaces, notably that the duals are vector lattices. Section 8 concludes the paper wdth a brief discussion of an alternative for- mulation, where the consumption space X+ the but rather all of X . Then is the uniform condition that distinguishes weak topology can be posed is T from as a local (topological) property. There are two companion papers to journal taken to be not the positive orthant this. Following immediately in Mas-Colell and Richard [1987], which does the hard work this of establish- ing existence of equilibrium for economics based on our (and other) topologies. Appearing subsequently, Hindy and Huang [1988] generalizes and extends our analysis to certainty, models models where there is uncertainty and temporal resolution of that un- and they consider the impact of our topology on some of the standard in the theory of finance. 2. THE CONSUMPTION SET AND CONSUMPTION SPACE Consider an economic agent who lives from time zero to time one. This agent can consume a consumption commodity at any time during his natural to require that the agent sumption commodity, ' But see section 8. ^ and life consume only nonnegative amounts that the accumulated consumption span. It is of the con- is finite. Thus a life-time consumption pattern can be represented by a nonnegative nondecreas- ing real-valued function x consumption from time zero consumption time to at "rates" (if x{t) x{t) is discontinuous), (if -+ $R, with [0,1] : t . denoting the accumulated x{t) Note that this representation permits absolutely continuous) and in discrete lumps is and more besides. As x is nondecreasing, it has the following properties: For each (i) (ii) t G [0,1], both unilateral limits x{t-) \im.3ii x{s) exist For each e t and are [0,1], x is The only possible kind and x{t+) limj^^ x(5) = finite. continuous at if t and only x{t) = x{t+). of discontinuity is a x{t-) (iii) = To avoid complications, we = shall further \\mx(s) = jump. assume Vie x(t), if that x right continuous: is [0,1). ait That is, the agent if commodity during time at f . consumes the time interval {t-,t+), he in fact To sum up, we ^ amount a nontrivial stock will represent the of the consumption consumes x{t+)-x{t-) agenf s consumption life-time pattern by a positive nondecreasing right continuous real-valued function x on [0, 1] . Tl-ie set of Throughout [r, 1] will That is, is denoted by A'+ the paper, Xr for r e [0, 1] be denoted by < . We of any x e defined. a, 6, denote the indicator function of for . Also, scalars ... will continue to X will . Xr represents a unit gulp of consumption at time r Here we have used the convention ^ V< . such functions use this that x{t) = x(l) Vf > 1 and x{t) = convention throughout, so that the value less than zero is always well arguments greater than one or X Let denote the linear span of Xj^ which , the space of right continu- is ous real-valued functions of .>ounded variation defined on > and > on induces orders and and > ii if - ^\ X^\ ^2 ^ {0} We . > ii X will refer to 12 The cone . ii if - 12 £ X^ ^+/ as the commodity space to Xj, as the agent's consumption set. Note x"*" 12 way: A" in the usual [0, 1] X is = x(0). . = Then we can define sup(x,y) X, we For x € a lattice, in the usual fashion. denote the positive (increasing) part of x to x+(0) that Note that - y^ + (x > x(0) if will use , then with inf(x,y) J// defined similarly. Our agent will be assumed to have preferences over Xj^ mathematically by a complete and transitive binary relation assume throughout X+ Vx e The reader may be put always will would suf- . xi-|-X2^Xi; that they are strictly increasing in the direction of xo assume We on A'+ that these preferences are increasing: Vxi,X2GX+, fice to >: that are given and Va > 0, x + ' axo >- x- played by xo here. off by the special role that there is some x e X+ in the direction of strictly increasing, although condition ness assumption about to be discussed (3) It which preference is below and our uniform proper- would need to be modified accordingly. For the analysis of Mas-Colell and Richard [1987], one would wish to use the social endowment Moreover, with T+ we in place of xo will assume denoting the topology • that there T . given a linear topology relativized to then be assumed to be continuous in T is . then preferences will T+ and uniformly proper with Continuity of preferences in a fixed topology justification. X+ And is fairly Uniform properness, on the other hand, is T on X; respect to standard and needs no not so standard. This condition, introduced by Mas-Colell [1986], lated to the existence of equilibria for commodity space and, indeed, is shown there to be intimately re- economies over an infinite optima to the existence of individual sumption budgeting problem when there are many infinitely dimensional in the con- commodities. The reader should consult Mas-Colell [1986] for further commentary; here, merely with respect Preferences scalar a > we — x , V ax* + av x '^ 3. cally. in properties will two Ideally, parts. T = form properness and we we will wish LIST for the want will a locally T V v e and , proper with respect to to T FOR T T ? The topology T to list of desiderata be well behaved mathemati- convex, metric topology, and A' should be this metric. will (eventually) for , • A WISH we should be X+ T- a ^ xo we want First, X^ and . complete and separable under agents, there exists a vector x* e if that preferences are uniformly for the particular choice of i* What are said to be uniformly proper of the origin such that, for every x e assume will X+ on >: T topology to a linear open neighborhood comes will recall his definition: Definition (Mas-Colell). In fact, we Also, since wish to we wish to invoke uni- undertake optimization be a linear topology. (That is, for our the vector space operations are continuous.) In order to apply the equilibrium existence results of Mas-Colell [1986], one with respect it to would wish T, or (even) will turn out, will be too that much that the lattice operations are continuous (A', to considerations. Happily, Mas-Colell X+,T) forms hope for, Banach lattice. But this, when we come upon economic and Richard To repeat: Our use of xo in this special Any X* e A'^. could be used instead. ^ a [1987] allow us to role is weaken this purely to ease exposition. desideratum substantially: and the topological dual to The second and more of the situation. to be Our and we close, We will want the order intervals to be weakly compact order dual. lattice in the be a vector substantial part of our wish X on concerns the economics consumption streams ought intuition suggests that certain will look for a topology list and X+ that accomodates this Consider the following requirements, phrased in terms of properties intuition. ^. of Two (1) patterns of consumption that have almost equal accumulated tion at every point in time should be sup lim close. Formally, \xn{t) — = x{t)\ consump- if 0, ""*°°t6[0,l] then x„ In (1), >: y for all t . fixed size across small we p(i,y) The idea of this tion path X is is We wish , = metric >: x„ for all n implies y amounts of time this are also regarded as insignificant. known metric on A^+ , To do the Prohorov metric. let inf{e is > : x(i + e) + depicted in figure e 1; > y{t) the e > x{t - e) - e}. neighborhood of a consump- y that lives entirely within a "sleeve" determined by moving e up and The x .^ so that shifts in consumption of a any other path (e in each direction) >i consumption between x^ and x must be weaken to begin by introducing a well For X and y in A'+ that , differences in accumulated small at every time so, n implies x >zy and y from x at to the left and down and every point {t,x{t)). Hence if y is around x to the right within e of model where consumption rates, unless felicity is linear. We will demonstrate this explicitly in section 6. The astute reader will also note that (1) could be said more compactly as: T'^ should be no stronger than the sup "* astute reader v^ll notice that this rules out the standard utility is the integral of felicity of norm topology relativized to A% . 8 > . X , then at every time consumption under the total consumption under x t , some time no more than y at is within away from e t of the total e With . we this can pose our second requirement: ^0 p(i„,x) If (2) y>zxn Our for final all — n as > CO, then x„ n implies y >: x >: y for amount shift of (2) somewhat. The idea of consumption across a small It may As n goes a positive scalar. grows We large. wish a^^d to infinity, the shift in time in particular, p(anXi/2ianX(n+i)/2n) —^0), but shifted in (2) amount Suppose is of time we that help to give a concrete example: Consider comparing the consumption patterns c„xi/2 is and an increasingly larger amount of consumption over an decreasingly smaller amount of time. On , ^ should not be of great consequence in terms of preference. imagine instead the y >z . economic consideration sharpens that the shift of a given n implies x all if a„X(n+i)/2n/ where grows small (and so, a„ grows with n, the amount to assert that, a priori, it should be possible for a„ to go to infinity slowly enough so that the difference in terms of preference becomes negligible uniformly. strict monotonicity of >- wish to assert that the "added shift of a„ from 1/2 To be more in the direction to (n utility" + l)/2;i,if precise, that xo we know, by xo in slow enough. Or we might require that a^ On is a„ >- OnXi/2 • We for the a^ goes to infinity sufficiently slowly. For we might if °nXi/2 from xo eventually compensates example, insist that, + presumed the o{n) is (i.e., anjn -^0), then the growth is o{n^l'^), or o(lnn), in order to come to the desired conclusion that the shift in consumption through time small relative to the is "sufficiently time is slow" rate of growth ^ of consumption shifted. But there in the amount shifted such that, if is some the shift in small, the change in utility is small. To measure what that amount we Or: is meant by a "sufficiently slow" rate of growth, we suppose are given a concave, continuous, strictly increasing function ^ T^ should be no stronger than the 9 weak topology on Xj^ : [0, oo) — such that [0,oo), = fi{z) z^Ip numbers a„ since (3) /i and n{n) — = + or ^{z) , ;z meant is as above, /z /i that the condition grows, so all , x^ "slow enough." We We are {j/n} + ^ xo we do two sequences of claims with and , (ii) = z . . fj. We is the appropriate function x' e X+ -worst element of A^(x) the worst the lower-right we It are imagining that ^ not wish to prespecify what rate /i is not very natural - for a is /x is given consumption path x, First, for a that are within is x "delayed consumption pattern boundary of the sleeve. 6, if some impatience is possess this path. What and lowered" if by, in is the e is: and lowered by in the sleeve If For each x G e ". what path x minimizes this we is, x in X+ We the would be to this structural property we suggest at the to have = k this . we can Suppose you path "delayed give you an additional xo all , terms of to that representation), then amount, over 10 That drawn around one admits added Nt{x) let of x in the Prohorov metric. most you are willing compensation how would one given population of consumers? ask of a given consumer: Take any path x € A'+ with i(l) Now then becomes stronger the observe, will hold for the sort of representation end of section < 1/n, p(y„,a:„) enters as part of the data of the it suggest that a natural property for consumer preferences (which, that are willing to contemplate, fixed /z(z) has been said to us that this condition 1, Note only suppose that some "slow enough" rate can be specified suggest the following. figure o(/x(n)). economic application. denote the set of >: sequence of Vn agents in the economy; consideration. But discover what we = a pose the following condition. when economy under It we depends on the strongest it is given uniformly for in a particular a„/fi(n) -^ 0, or a„ {x„} and if For examples, think of . By "slow growth" of . growth" that In general, for all sufficiently large n faster 1) and {y„(l)} are both o{^{n)) {a:„(l)} Note oo as n -> oo > ln(2 the fastest "slow is must be concave. For fixed (i) 1 relative to growth linear p > for = //(O) paths with x(l) ^P front? = kl (For preferences with representations as at the end of section problem answer is tractible.) Finally, how does this amount change do these in terms of optimization as k increases? The THREE CANDIDATE TOPOLOGIES Several topologies on the space of functions of by mathematicians. In extensively analyzed this us fi.^ to this question, of course, gives 4. 6, our vdsh bounded variation have been we this section, ask how each of list. For the space of functions of bounded variation, a mathematically natural norm is total variation: For x e X , the total variation of x , denoted by TV(a-), is TV{x) = |x(0)| + / \dx{t)\. Jo Note well that positive cone, variation if ' norm to have X be will need to have we hope then we it lacks separability), it wish In the first place, (1) is x„ given by Xn(t) According to (i), = k/n for t the underlying time space Although at least as violated: Consider the € {(k - strong as the total is consumption patterns l)/n,k/n] and x given by x{t) Xn converging to x, but (of course) norm. In the same way have the sort of economic to topology performs miserably in terms of our economic we should have not, in the total variation when a with A'+ the topology. But while this topology performs well mathematically (although list. a topological vector lattice a (2) is violated. continuum, measures time qualities we = t. we do Total variation, much too finely desire. getting a bit ahead of the game, let us observe that there will be a point to these questions. The "faster" the function ^ rises, the "more continuous" we expect equilibrium prices to be. This is not quite precise logically - our characterization of equilibrium prices as coming from the topological order ^ it is dual will depend on uniform properness in norm topologies that are motivated by but not implied by (3). But, in spirit, the idea is that the "smoother" are consumer preferences, the more tightly we can characterize prices. ^ which we would want if we were to rely on the existence result of Mas Colell [1986] 11 . Consider next the Skorohod metric originally introduced on the space of functions which are right metric atinuous and have Let defined as follows:* is continuous mappings of For i,y G which c X, onto [0, 1] A denote A = = and A(l) infimum of those positive define S(i,y) to be the there exists in the class of strictly increasing, A G A, then A(0) itself. If (RCLL). This finite left limits scalars 1 for e such that a A sup \\(t)~i\ < e <6[0,1] and sup - \x{\{t)) y{t)\ < e. te[o,i] It can be shown that S : XxX -^ R.^ is a metric, and that {X, S) a separable is metric space. Moreover, this metric does possess two consumption patterns desire. Since under the Skorohod metric, note, in this regard, that 2n{t - 1/2) for p{^n, X1/2) -^ t € if [1/2, (n (so (2) some + define x^ l)/2n], by comes ^ t , it is for for 1 t > not satisfied. < f (n We 1/2, x„(i) + = l)/2n, then should apply), but !„ does not approach xi/2 in the is also not very well to exploiting the linear structure of the In particular, tn = (2) is = Xnit) and x„{t) Skorohod topology. The Skorohod topology it Unhappily, we sup norm are close that are close in the (1) is satisfied. we of the economic qualities that behaved when space of consumption claims. not a linear topology. Whereas X(„ -^ Xt in this topology their difference x<„ - Xi does not approach zero. Hence we if continue our search. One other topology frequently used vergence denoted by T^ C[0, 1] ® A , on X is the topology of with => denoting convergence in be the space of continuous functions on good reference is Billingsley (1968). 12 [0, 1] . weak this topology. Then Xq =^ x if conLet and only . / f{t)dxcit)-^ I on the closed (Integrals here are defined as a first temn in the integral.) separable It is (cf. Holmes interval [0, 1] , so that /(O)x(O) enters ^ weak topology Recall that the V/€C[0,1]. f{t)dx{t) Jq Jo a linear, Hausdorff, locally is [1975, exercise 22c]). convex topology. metrizable, It is by precisely the Prohorov metric p (hence has a countable neighborhood base and , on A'+ , but From not metrizable on it is the point of on Altogether, all of A' ). It it is . weak topology on in the .Y+ will obey quite an attractive candidate (except that has the further virtue analyzed in the economics (in is terms of in (3), — 2^/TJJi(lJt) linear preferences by x ^ y + if 1. for ji , weak topology and consider the continuous / fdx > J fdy . That = oo is weak is assumed.) Define preferences are given by is, evaluating consumption plans at the "prices" given hy f because / and any fixed function ^, however. (Recall that //(oo) tinuous and uniformly proper in the (1) Mas-Colell [1975] and Jones [1984]. some so even for linear preferences: Fix function f{t) well consequence) of having been thoroughly literature, e.g., in need not be well behaved is not metrizable it is Preferences that are continuous and uniformly proper in the This sequential) view of economic desiderata, the weak topology behaved: Preferences continuous (2). X is topology, and . it A fortiori, y is strictly is con- increasing strictly positive. Consider x„ qualify in terms of = \/M^Xo (3). = and y„ \/A'(")X'i/n In particular, x„(l)//i(n) = • Note 1/ y/yjn) — that » . a:„ and rjn But, for every n, / < fd[Xn j fdyn + Xo) = Va<^ = • 1 (1 + a/M^ + 2VlM"))- ^ If we norm C[0, 1] with the sup norm, then the weak topology topology on A' relative to C[0, 1] 13 is the weak- * . Roughly put, "shadow prices" consumption for time zero rise at times just after very quickly - so that shifting an increasing amount of consumption from time zero to time 1/n has quite an impact, Condition so that the holds that there (3) is a if should become negligible, but the shift to rise increase is still we have too much. ^° A FAMILY OF NORM TOPOLOGIES total variation is topology does better, but convergence in the family of is violates (2) this, we and not is linear. violate (3). In this section, topologies which will be free of the reader is must be introduced an accessible reference.) all The topology . The Skorohod The topology of weak quite suitable, except that preferences continuous (Musielak [1983] contained, it X too strong along the time dimension. weak topology may norm To do slow- picked shadow prices (given have discussed three possible topological structures on generated by proper p>oint of (3) is that the around zero quickly enough so that the prespecified slow-enough- 5. We increases quickly enough. slow enough increase in the amount shifted enough-increase must be prespecified, and by / ) amount that we and uniformly will introduce a these defects. to the notion of To make this an Orlicz space. paper close to self- give a very quick introduction here, adequate for (and told in terms oO our particular application. To begin, consider the case where the function ^[z) Imagine that we defined the following norm on X = z^^p , for some p > 1 : i/p l^llp = ''dt + \x{l)\P .Jo ^° For any continuous function /, and for preferences defined from / as in we can find a rate function /i that grows slowly enough so that (3) would be satisfied for these preferences and for Thus the order of quantifiers in (3) is all important - ;i must be picked first, and then, what we have shown, there are "nice" (i.e., linear) preferences continuous and uniformly proper in the weak topology that violate (3) for this fi. the example, f.i 14 . Since our functions x are tegrals exist. of bounded variation, From standard Banach space normed separable all theory, it is clear that all these in- clear that (X, it is • || ||p) is a topological linear space: Nonnegativity of the norm, the triangle inequality, and the standard conditions in conjunction with the linear space operations all X standard theory. Right continuity of the functions in = then 1 follow immediately from ensure that if = ||a:||p 0, 0; note in this regard the special treatment of x(l). For separability, note that simple right continuous functions that change values at (finitely many) arguments and take on rational values are dense rational with this norm. norm: Consider, ation functions 2),l/(2u - 3)) worth noting, perhaps, It is for ... U ( A+ , however, This has the right "flavor" for • where if lip II (3) is to the difference in amounts as well, for the case n{z) are scaled by being raised two consumption paths What, then, do we do if y.{z) obvious candidate analogy is to , ft T]{oo) Tj to ^^ T] = continuous, is oo . in this is is 0(l/n) begin by defining [l/(2n - obvious tj to the See proposition 4 for an exact statement. . in Roughly amounts, power p Hence . dimension and small. ^^ z^/'p ? For a general is — fi~^ . From is, p. , an the properties 7](0) = value repeatedly, extend real line; that 15 U z^^p and convex, and satisfies to write absolute be an even functions on the entire 1)) to the = in the time not of the form strictiy increasing, To avoid having - complete.) o(n}IP) in the space dimension, then the difference of X in has unbounded variation (and trades off differences in consumption paths in time differences in hence complete These are Cauchy and converge [1/2,1]. not right-continuous). put, , example, the sequence of right continuous, bounded vari- measurable function, but what they converge is L^ that A' will not be that are indicator functior\s of [l/2n,l/(2n a:„ U in for negative and jj. and arguments X , = let \i{x) and ^l{-x) let 7/(1) = lklU:=A^(^ Unhappily, this = A r]{x{t))dt Hence norm, r]{x{l))y is a homogenous function, ^ the triangle inequality will not follow, nor will more ingenious procedure define the unit + for the try, for scalars a. |a|||x||^ slightly Then . won't work, because while z^^f will not be, in general. ||ax||^ r?(-x) X circle in , C, := {x denoted eX: f C^, is as follows. For ^ and as above, 77 as , rj{x{t))dt + r^{x{l)) = 1}. Jo The convexity of "gauge" to for ensures that this unit 77 measuring all other vectors x be the unique positive scalar a that because of the continuity of this definition is 77, ||x||^ is circle is : appropriately shaped to be a Define, for x other than x satisfies x/a ^ It Cr, = 0, ||x||^ takes a proof, but well defined. Moreover, for = fi{z) z^^^ With equivalent to the standard definition of the norm. , this definition: Proposition (A', 1. • |1 ||^) is a separable normed topological linear space. the order operations given above, order intervals are weakly compact With each of in these topologies. We provide the proof in the appendix for completeness, although application of the theory of Orlicz spaces, of which this Musielak T;, , simple an example. (See [1983].) Fixing by is this is a /i and the corresponding and the the case ^(z) = = z the topology induced topology relativized to T,, case where fi{z) 77 , z'^Ip , we A'-i- is denoted by use notation such as Jp , by T^ results are established in the appendix. 16 . is denoted For the special and we use Too . The following basic ||-||^ / etc., for Proposition 2. For turn The is any ^ and corresponding weaker than both the Skorohod and first part of this proposition Ti r/, is weaker than T, which , in total variation topologies. Since is fairly intuitive. t] is convex, tends to it accentuate (more than linearly) large differences in the dimension of space. Note "weaker" in that also that, in our this proposition norm should be read as "no stronger than." Note - continuous. For example, xi/2 >: that is looking for preferences on that >: is to T^ , increasing as before.) Proposition 3. Preferences (2). that are given A'^. some prespedfied for The following for the "right" thing, insofar as and do not converge. T+ complete, transitive, continuous in the proper with respect Moreover, .Y+ is t] binary relation a topology, and uniformly (We . by will continue to results establish that our desiderata were that are continuous >: definitely not X{n+i)/2n -^ 0/ but the corresponding positive parts (or negative parts, or absolute values) We will be most topologies, the lattice operations are T^ separable under we are looking T^ satisfy (1) correct. with respect , to and hence every continuous y on X+ admits (complete and transitive) preference relation assume a continuous numerical representation. Satisfaction of (2) follows as a simple corollary to proposition 5 (which proved (1). The in the appendix), separability of which X^ immediately implies the satisfaction of follows standard lines and of a numerical representation for in turn is is is omitted. The existence standard given continuity and separability - see, example, Fishburn [1970]. Proposition 4. Fix /i and the corresponding ?/ . Suppose that y is T^ continu- ous, increasing as above, and T^ uniformly proper /or the specific choice i* Then y satisfies (3). 17 = xo • we prove one Abbreviated proof. In the appendix, as in - x„ then y„ (3), -4 yn — Xn is eventually inside Xo + ^n, to conclude X in the relation T,, . . Thus V Apply to xo + a;„ How This leaves open one issue: weak topology on X+ ? It is clear to the Proposition 5. weak For any topology on Xj^ and y„ are of the origin, y„ must not stand it. do the topologies T^ compare with the from the example of section 4 and proposition 4 that uniform properness with respect to with respect V any neighborhood xo+^n- Xo + J/n - 2:„ = That does . If j:„ the definition of uniform properness to . that, evenhaally, for technical fact: T^, is stronger than uniform properness topology. But: and /i T^ t] , is topologically equivalent to the weak • The proof of proposition 5 is essentially technical, and so it too is consigned to the appendix. Hence we Uniform properness reaffirm (the obvious): is not a topological property 6. ON THE STANDARD MODEL AND THE The standard model T^ TOPOLOGIES of preferences for consumption through time assumes that preferences are defined over consumption in rates and are given by a nu- merical representation of the form U{x)= u{x'{t),t)dt, / Jo where x' denotes the one specializes first further, to u{x'{t)^t) ing function, although for continuous derivative of accumulated consumption x in its = . Typically, e""^ f{x'{t)) with /(•) a concave, increas- what follows it two arguments. IS will suffice to assume that u is jointly , Preferences so given are not defined for consumption in gulps (where x we might discontinuous), but preferences to general x approximate x : consider the following procedure for extending Find a sequence of absolutely continuous x„ = and then define U{x) , is lim„_oo U{,^n) > hoping that that the limit well defined and independent of the approximating sequence x„ . We will is want preferences to be T^+ continuous, so the sense of approximation had better be no weaker than to convergence - and T^ approximate continuous x^ we have arbitrarily closely (in . Hence already remarked that terms of T^+ any x G ) program does have some hope this A'-j. it is possible by absolutely of working. It will only work, though, in very special cases: Proposition (a) 6. Suppose that preferences on by are represented A'^. U{x)= :)= / / the utility function u{x'{t),t)dt u{x'it), (*) Jo for some (jointly) continuous function u sumption paths X. to absolutely Conversely, (for some if, 77 ), for all absolutely continuous con- these preferences are continuous in continuous consumption paths the form u{r^i) (b) If , — OL{i)r + (for any where a and ^ are /?(<), for a preference relation >: on A'+ t] ), T^ , relativized then u must have (clearly) continuous. that is continuous in 7^ preferences for absolutely continuous paths can be represented by [7(x)= / a{t)x'{t) + /3{t)dt Jo for continuous functions a and U{x) /3 then , = f a{t)dx{t) Jo represents preferences on all of A'^. 19 . And for M any continuous function a preferences on Proof. Suppose as in (+), r for and some concave in is not r', {u{r,t)+u{r',t))/2 its first This . continuous for the is Then affine. for 12 all on absolutely continuous paths u(-,i) t, <u{(r + r')/2,t) u{r',t))/2 T^ given hy i*-k) are that preferences and arguments X+ are represented some (nonnegative) u{(r+r')/2,t). Suppose that {u{r,t)+ ^ more natural argument. But the other case case, thinking of u as being handled by a virtually identical is argument. Then by the continuity of around positive length — r')/2,t') t such that for Consider e.^^ u, there exists a all t' > e el and some {u{r,t') , + construct x„ as follows: subdivide , consume at rate r at rate r on I = 1 the even subintervals for a bit of time, then at for < ^ J and + U{x) > U{xn) x'{t) < u{r',t'))/2 u{{r + r' , = and is each on the odd. (That r' and then (r+r')/2 for eA(J), where A(J) in 1 a-„ . into 2n equal sized subintervals, consumption paths converges weakly (hence x'{t) of sequence of (absolutely continuous) consumption paths constructed as follows: Off of I, consume at rate I J interval in t at . ) and consume This sequence of etc.) , T^ any eJ r is, On to the path x that has Hence, by the representation, the length of the interval J. It is easy to turn this into a contradiction of the supposed continuity of preferences. To is finish part (a), note that if u(r,f) jointly continuous, then For part (b), lutely continuous are of this all dense is that, of (in A'^. (^Hk-) , a{t)r ^{t) for r £ (0,oo) if u functions. T^ and the consumption paths which are abso- any of these topologies) except and gives a weakly continuous (hence when "felicity of constant marginal felicity at every point in time '^ + must be continuous fS note simply that continuous) function on The point a and = (a . consumption rates" exhibits case not of great interest in the This does not assert that preferences so defined are See the discussion at the end of section 7. '"^ The cases t = and t = 1 present no difficulty 20 in A'+ T^ uniformly proper. standard model literature), the in the example given norm tof)ology, so be continuous in our topologies. Indeed, converge to x in the sup in the proof of part (a), the a:„ any treatment of preferences that asked for property (1) would not admit The reason why should be ing steadily fails to the standard Unless marginal clear: consumption through time for model in the literature. felicity is constant, consum- not well approximated in the standard model by consumption in is "quickly" and properly spaced gulps or even by consumption at quickly varying in the sort of model might then be used Following Jones (1984, p.525), we make problem with the standard model at time t of felicity for some time some length 5 that continuous 6 [0, 1 the + : One [—k,0] k] — > standard model? to replace the The the following tentative proposal: consumption from at time generates t consumption taken in at time way of time thereafter - put the other at time t felicity A more gone immediately. it is depends on consumption not only in the recent past. G is only; the "pleasure" derived model would be one where ^ close sup norm. What for made cumulative consumption paths of such patterns can be rates. Yet the produces around, a apt flow felicity at but on consumption s can imagine a model of this running as follows: For [0,oo) and consumption pattern x 6 by Fx{t) = /°^ eis)dx{t + 5) ^^ That . is, A'^. Fx{t) define Fx{t) , is a "weighted average" of recent past consumption, where 6 gives the weighting function. Then apply the usual construction of utility from felicity by defining /•i+fc U{x)= for some u{Fx{t))dt, / Jo well behaved "instantaneous felicity function" u the reader, but one can generate conditions gives ^'^ T uniformly proper preferences. Since this is only an example, Note that ^(0) = will 6{0) = to savour a jolt of consumption. . on u and We leave details to 6 so that U so defined (Indeed, one can generalize this quite make matters come to mean 21 . easy by assuming that 0{-k) — that it takes some time to begin weighting function 6 can depend on calender time, substantially, so that the Since etc.) we have no particular reason preferences (except perhaps for this model further. its to advocate mathematical this specific representation of tractibility), we wall not pursue But the form given does indicate the sort of "intertemporal substitution of consumption at near-by times" that motivated this study. Indeed, need it has been said to us that this form That for the analysis here. is (to is so natural that it obviates the quote from a referee report), "one can think of the actual consumption process as the process of 'burning' calories, of transaction costs associated with eating." for that how We agree with actually motivates our study. That it food is The discreteness of the eating process occurs presumably because quite smooth. beUeve which is is, if this one has statement, but this picture in we mind translated into calories or satisfaction, then the sorts of topologies we have advanced (or the weak topology) become more natural than the standard model. The payoff then comes from the resulting characterization of equilibrium prices; if preferences are "smoother" (continuous in a weaker topology), then prices (as continuous linear functional) are taken So we turn now from a more to the issue of characterizing (positive and) functionals for our norm continuous linear topologies. DUALITY 7. For the duration of restricted class. this section, we assume that some pair ^ and 7/ as in Section 5 has been fixed. The companion paper by Mas-Colell and Richard be possible to establish existence of are increasing [1987] tells us that it will equilibrium in economies where preferences and uniformly proper in our topology if we can establish that the topological dual forms a vector lattice in the order dual. Moreover, equilibrium prices for these equilibria will be our aim drawn from the topological order dual. in this section is to characterize the topological which we denote by X* , and to verify that A'* 22 Hence dual space of (X,T;,), does form a vector lattice. We will not give (e.g., complete answers here, as the theory in Musielak, 1983, section 13). But we well developed elsewhere is will give some reader can see the basic flow of the argument. In particular, dual to (A',T), denoted A'*, (essentially) the is of details, so that the we will show that the space of Lipschitz continuous functions. As a first step, we show that every continuous linear function can sented as integration against some continuous function on is no larger than And C[0,1].) [0, 1] be repre- (That . is. A'* the order dual will correspond to nonnegative continuous functions. Proposition 7. l( (f) e A'* , then there some / e is C[0, 1] such that = f mdx{t). <f>{x) (0) Jo Moreover, the positive cone in A'* will correspond to continuous functions / that are nonnegative. (Recall that Proof. (^ when we write J^ , we mean The proof follows standard € A'* . For t Xt„ —* Xt in T^ that that, by [0,1], define E , we know linearity, function x € A'+ And by (0) on A'+ . proposition is trivial. <f) f{t) that / = is and so (l>{xt) we only sketch Since t„ -^ t a continuous function in [0,1] on [0, 1] a sequence of step functions, to obtain that then extend to If / gives a all some implies that . Next show , of X by linearity. The final = > f{t) f fdx > 0. Conversely, if / statement of the in section 4 23 (<>), is and / is nonnegative, . course, not every continuous function / gives rise to a example given has the form <f) continuous linear functional via J fdxt then for any nondecreasing x linear functional - the First fix it. has the form (0) on step functions. Approximate any in the positive orthant, then Of lines, the integral over the closed interval.) shows this. T^, continuous This example also . shows the fundamental thing that can because, while continuous, continuous linear function, a relative to /i go wrong - the function / may it varies "too quickly" r e [0, 1] , define n n = sup{^ - \f{t,) f{s,)\ < : 51 < <i < ... < < 5„ < <„ is, Vf{r) (i) is maximum the is (ii) < r). absolute variation of / over a disjoint finite whose lengths sum to r or less. Since / is continuous defined on a compact interval, the following facts are easily established: The function / has unbounded variation over if s,) i=l collection of subintervals and - Y^{t, 1, 1=1 That give over short time intervals. For a given continuous function / and scalar Vf{r) fail to = Vf{r) oo for The function / r all > the interval if [0, 1] and only . and only is absolutely continuous on [0,1] is the definition of absolute continuity if if limr-^o ^/(r) = 0. The second of these The is first does have a = that Vf{l) oo The point of condition on / Proposition according A 8. . bit of content, since the definition of But / is to unbounded variation not hard to get the stronger statement. with we it, can give a simple necessary give a T^ continuous function. continuous function / gives a T^ continuous linear functional (0) only to it is this definition is that, if - no proof is necessary. if there is K some constant such that ^{\/r)vj{r) < K for all r Proof. Suppose ^(l/r)v/(r) Si) < T = N . Then we can and /i(l/r)^, or on the xt^{t) > to the contrary that for r? |/(i!,) - find ^i /(-^i)! - f{t,)) X fi{l/r) < > ii A'- A'^ < . Note why 24 there exists ... < 5n < t„ some r such that such that ^.(^ - Construct an element x;v within 1^(0 = unit circle as follows: sign(/(5,) every for < it is G [<„,5n); for t 6 [5„,^n), that x/v lives within the r/ s unit circle: which = a:A'(l) — r]{x^'{t)) lem.) on . This 1. only Corollary the linear functional given unit circle. T] Corollary (<>) J^ f{t)dxf^{t). is, if Then / € measure r or less, on It ti = |/(i,)- does not present a prob- s,+i by / and (0) takes on unbounded values cannot, therefore, be T, continuous. A function / G C[0,1] gives a T^ continuous linear functional via / is absolutely continuous. Suppose 2. set of by construction, precisely /^(l/r)X^ (The reader should check that the case Thus the nonzero on a is l/r. Now evaluate f{si)\ 0, and x^it) that is rj asymptotically linear; rj{r)/r —y k for some C[0,1] gives a T^ continuous linear functional via (0) only ^' if > . / is Lipschitz continuous. Corollary 1 follows immediately the second fact given above and the proposition (and the fact that /^(1/r) -+00 as if 77 we is asymptotically linear, then so of the order 0{l/{ti^ Corollary /' ^ is For corollary 0). > . And if / is 2, use the fact that not Lipschitzian, then can find a sequence of intervals [5^,<;v] of vanishing small length such that \f{tN)- f{sN)\/{tN-SN) > is — r 1 - N sim)) , . But then and our f^{l / {t criterion will provides a further harvest. differentiate almost everywhere and as the derivative of /, / we is is f{t)dx{t) = Proposition 9. If a / is be violated. absolutely continuous, then the integral of /(l)x(l) its it derivative. Writing - / x{t)f'{t)dt. (00) Jo included in the first integral, from other applications of integration by which includes If will be (asymptotically) can integrate by parts in (0) to obtain ^0 (Note that /(O)x(O) n - n)) converse to corollary / € C[0, 1] is parts.) so this may look a bit different Hence we obtain the following, 2: Lipschitz continuous, then / G A^* for every 26 tj . / Proof. If K constant Lipschitz continuous, then is Hence . for any x on the rj{xil))+ K < |/'| some unit circle, since rj Tj{x{t))dt f (almost everywhere) for = 1. Jo we can use the convexity of and thus t] to conclude that 1 and that < T]{x{l)) x(t)dA < r]( 1. I Using (00), us an immediate upper bound on this gives A")^(l) for X from the is bounded, and thus unit circle. tj Hence X* where is well that = ri{r) r worked . r]*{s) = sup{r5 — r]*{s) = 5', function, the norm of r -+ as T]{r) r : it is • || ||^. ; i.e., for for If, much more complete char- a 1 . Then / for K /' , Assume T]{r) — What if r^ to give a for the for 1 moment ri*{f'/K) T]{r)/r 7^ > p > rj* 00 , by then T^ continuous linear live in the Orlicz development can be found general setting). The theory the "conjugate" function rj example, some constant (in a space given by must be integrable and in Musielak, chapter 13 as r —> 0? If 77 is not which case we know the answer from proposition 9 above), then we can create an past . necessary and sufficient that integrate to one. This 77 0} where 1/p + I/5 = must linear (in Define from 0. > we have that are not asymptotically linear? 77 and runs roughly as follows: out, —^0 ri{r)/r 2, asympototically linear, including the special case is 77 What by (0) continuous. it is if f{t)dx{t) of (sup|/(<)|+ the linear functional defined Putting together proposition 9 and corollary acterization of J^ some r]' which does meet level r . The proof depends only on the behavior of this requirement and which of proposition 2 in the 77 for large 26 is identical with appendix shows that T^ arguments, so the dual with respect , . to T] will be the same as with respect We Musielak. to and the r/' , leave this to the reader, only to note the special case of to state that the dual space a lattice in the order dual, is by latter is characterized and to T](r) make two = r^ further remarks. Proposition 10. If = ri{r) rP for some 1 for all absolutely continuous functions (<0>) for q satisfying l/p + = l/q any Proposition 11. For the space A'* rj , nonnegative functions / (for the X* f,g e oo, then given by the form /' lie in L', Hence . is a lattice in the order dual. cone defining the order dual corresponds in the order dual, the join of / and g obvious embedding) would be given by max{/, y}, and in the of the derivative of max{/, ^} less is / whose derivatives meet by min{/, ^}, these being taken pointudse. everywhere) X* 1 Proof. Proposition 7 establishes that the to <p< than |/'| + for absolutely \g'\. The Now the absolute value continuous / and g rest is clear is (almost from the characterization of A'* given above. Remark we 1. Lipschitz continuity of / 2. = Since U 77 . proper it is In proposition 6, a, U{x) /o is a{t)dx{t) that, framework of are "within" the usual Remark show Propositions 9 and 10 the we same once we the classic as /' G L°^ are able to integrate Banach spaces; by parts, in particular, . asserted that for any continuous function of time would give preferences that are T^ continuous for all (manifestly) linear, for preferences of this form to be T^ uniformly would be necessary for q e X* ; we now know the conditions needed for that to be true. 8. To this point, PREFERENCES DEFINED we have ON A' taken an approach where preferences are defined only over nondecreasing consumption paths x G 27 A+ . In some treatments of infinite , dimensional commodity spaces (e.g., be an entire linear space. In to defined on (although on [1981]), the this case, this consumption we could no rely so heavily on the of X ,\n proofs). all X on some open subset of (or, at least, set is taken would mean having preferences weak that contains somewhat taken this approach, our exposition would be Had we able X of all Kreps topology, which In place of requirement (3), we X+ .) less tortured is not metriz- could pose something like: For any sequence (3') 3;+(l) = > then X mean o{^{n)) y it (or ing in and , and for y if >: we want to + 2;„ that (i) x for n, then y all if Xn-\- >: ^ in the sense of (3') , we are discussion at Clearly, (3') all; on all is is (strictly) There of ^^ A ; A ). weak no need topology. Property set d(x,y) (3') = to (ii) for all y we fails in to the weak our norm invoke uniform propemess in the And we know all (strictly) norm less requir- ^^ weak smaller than the dual of the (Of course, the and (By x~ here, we move of A' there are sequences convergent in our whether our T^ topologies are all is a local property. are not convergent in the x>zy more or This to zero. topology (as the example given before indicates), and so this. • \/n .) such sequences {x„} converge topologies to ensure = p(x+,x~) choose a topology such that preferences continuous more) are well behaved (3') such that .Y every x and y from A', the negative part of x^ Essentially, since in , from {x„} topology. T^, topologies that that the dual space A'* One wonders, therefore, weaker than the weak topology (on topologies are not the weakest topologies that may suggest to the reader another metric topology on all of — y)'^ (x — y)~ This does not give a norm, clearly, and we p{{x ) , are not certain that it gives a metric: the triangle inequality is not at all obvious. But if it is a metric, it gives a topology that seems too weak to us. In particular, would imply properties like (3) without any restrictions at all on the rate of growth of the Xn and y„ And it would give very uninteresting equilibrium prices: The only continuous linear functional on A' that is continuous in the it . sense that ^(x„) -> if J(x„,0) -> is the trivial 28 (?i>(x) = 0. give A'* for the dual.) Our guess counterexample nor proof. is that this not true, but is we can offer neither ^^ REFERENCES Kenneth J. Arrow and Mordechai Kurz and Optimal Fiscal Policy, (1970), Public Investment, the Rate of Return, Johns Hopkins Press, Baltimore. Patrick Billingsley (1968), Convergence of Probability Measures, John Wiley New and Sons, York. Peter Fishburn (1970), Utility Theory for Decision Making, John Wiley and Sons, New York. Ayman Hindy and Chi-fu Huang Continuous Time Dimension, Richard B. Holmes II: (1988), "On Intertemporal Preferences with a Models with Uncertainty", mimeo. (1975), Geometric Functional Analysis and its Applications, Springer- Verlag, Berlin. Larry Jones (1984), "A Competitive Model of Product Differentiation", Economet- rica 52, 507-530. David Kreps (1981), "Arbitrage and Equilibrium in Economies with Many Commodities", Journal of Mathematical Economics 8, 15-35. Andreu Mas Colell (1975), "A Model of Equilibrium with Differentiated modities", Journal of Mathematical Economics Andreu Mas ical Colell (1986), Infinitely 2, Com- 263-296. "The Price Equilibrium Existence Problem in Topolog- Vector Lattices", Econometrica 54, 1039-1053. Andrew Mas and Scott Richard (1987), "A Vector Lattices", mimeo. Colell of Equilibria in New Approach to the Existence Julian Musielak (1983), Orlicz Spaces and Modular Spaces, Springer- Verlag, Berlin. ^^ A referee suggests the following counterexample, for the case where con- sumption takes place not on the time interval [0,1] but instead on [0,oo). For be the bounded variation function that rises each positive integer TV, let i^ to 1/A^ at times l,3,...,2iV - 1 and that falls back to zero at ttmes 2,...,2A^. Then x^ converges to zero in the weak topology, but it does not do so in any of our norms. We are not certain how to adapt this example to our setting with a bounded time interval. 29 Scott Richard and William Zame Utility Functions", H. L. (1986), "Proper Preferences and Quasiconcave forthcom'ng in the Journal of Mathematical Economics. Royden (1968), Real Analysis, Macmillan, London. 30 : . APPENDIX Fix a continuous, strictly increasing with 7/(0) = and = 7;(oo) and convex function and for every x Lemma Al. For x G,with in We a ^ + lima-.oo '^t;!^:,^) and that is ||x||^ appendix, we = linria^o part 1, and) only (if that scalar = -'^r/l^jCi) = cxd . such that . identically zero. triangle inequality, is is = ||6x|| • || .) || Of course, y{t) \ ( + x{\) / J^^ ||y|| for scalar + Hlkll + + ||0|| , is so that evident. For the h is 77 / ||x||(x(0/||x||)+||y||(y(0/||y||) \ l|xii we have apparent that for every x that |6|||x|| Hlkll + lly|ir H just defined, I^ (For the remainder of the 1. and write 77 it is use the convexity of \( x{t) + r^ right a The equation we is unique scalar a > such that I^{x,a) = 1 well defined, and it is zero only at the function x not identically zero, there our candidate norm = Ijj{x,a) From Lemma Al, Because x 0. In terms of the function 1. a x if suppress the dependence on defined to be Vo decreasing and continuous is strictly leave the details to the reader with one observation: Proof of Proposition is = certainly well defined is shown. easily is by T]{x{l)/a). variation, the integral 0, the function I^(x,a) continuous, I,j{x,a) is bounded Moreover, the following . [0,oo) to [0,oo) 00. Define /^(x,a), for x E A' and a G (0, 00), T](x{t)/a)dt Since functions x are of from 77 y{l) \_ I|y|i;~ ||x||(x(l)/||x||) ^i^ ||,|| + + ||y||(y(l)/||y| ||y|| <_iL + |2:|| ||y| Nl Ikli Hence Next, ||x + y|| < ||x|| we must show a useful lemma (now + + XI + ,11^1 „ lkll l|yll + , „ X 1 ^ 1. liy| ||y|| We separability. that we know pause that space). 31 for a we have a moment in order to state normed topological linear Lemma A2. A sequence {x„} every scalar a>0, X in lim„ /;,(in,a) = converges to zero in if • || 1| and only for if, 0. Suppose that, for every scalar a > 0, lim„ Iji{xn,a) = 0. Then for every scalar a > and N{a) sufficiently large so that I^{xn,a) < 1 for all n > N{a), we know that ||x„|| < a for all n > N{a). Thus ||x„|| -> 0, and i„ -> 0. (i.e., ||x„|| -> 0). Fix a > 0, suppose that n is Conversely, suppose x„ ^ sufficiently large so that ||x„|| < a, and write Proof. + Ir,{x„,a)^ I T}{xnit)/a)dt r){x(l)/a) Jo a where \\x m\,,^^f^il) Hikn <feJx a a \ \\Xn\\ / Jo the inequality follows from the convexity of \X, xl, ry . The right hand side clearly has limit zero. Proof of proposition 1, change values at that most Q denoted continuous functions set of all right finitely times, at rational rational values only. This set, dense Take the completed. is countable. As is well known, Q under the L^ norm. Moreover, in the set of all integrable funcrions, any fixed function of bounded variation x arguments, and that take on , we is for can find a sequence of functions Q which approximate the given function in the L^ norm, which differ from X in the sup norm by at most the total variation of x plus one, and that take on values at t = 1 that have as limit x(l) From this, it is easy to see that for a given function x £ X and a given integer n we can pick some t/n € Q such that Ij,{x — yn, 1/n) < l/n Apply lemma A2 to conclude that y„ — x from . , > . . To complete the proof of proposition 1, we need to show that order intervals are weakly compact. In fact, we can show sequential compactness in the norm topology itself. The proof follows standard lines, and so we will provide only a sketch. For X* this is > X, let / be the order interval without loss of generality because any countable dense set of points in [0, [x.,x']. all 32 will assume that x, of our topologies are linear. Let 1) that of X*. We excludes all = Q 0, be points of discontinuity . Now X e I and only ii x if is . — nondecreasing and x* x is nondecreasing. Let {xn} be any sequence out of 7. By the usual diagonalLzation procedure, we can produce a function y such that, along a subsequence (which we will hereafter = not bother to mention), limnXn(i) Xn nondecreasing, so is at all arguments t e Q- Since each — Define a function x by x{t) y. is y{t) for limsity{t) G [0, 1), where the limit is taken over sequences converging to t from (strictly) above along the set Q and x(l) = limna^n(l)- Since y is nondecreasing, so is < Moreover, y is right continuous (and has left limits) by construction. Clearly lim„ Xn(<) = x(^) at all points t that are not points of discontinuity of x - that X . for almost every is, t And . = x(l) by construction. Since all the x^ we can apply dominated convergence and lim„ x„(l) and X are bounded above by x*(l), lemma A2 We are to conclude that x„ approaches x in T^ done once we show > that x* > i We . have already remarked that x > 0). So let us carefully sketch the other half: If x*{t) — x{t) < x*(t') — x{t') for t > t' then right continuity of both functions implies that we can find values q> q' from Q such that x*(g) - x{q) < x*{q') - x{q') (The case oi t — 1 is left to the reader.) By increasing q' slightly to be some q" < q if need be (if 5' is a point of discontinuity of x this gives x*{q) — y{q) < x*{q") — is nondecreasing (i.e., x , . , > y{q") for q Which q" Hence . We note first that for any 77, T^ is topologically equivalent any scalar a Hence, without loss of generality, we can assume that Note that this implies, by the convexity of rj that ri{r) > r for all to Tajj for r T] > = 1 , Now it is 77(1) = 1 (with maximum any t] . of any only need we ) Xn to — let two convex T^ rj , If evident that The proof of for A2: 2. . 1. . is Tr, that , the limit exceeds J sum ||^ first if for that part of proposirion 2 x„ for ||^' || we have Tr,^ — > . Of course, > rj > T is is integer > . Now if convex as is shown T^, — M Hence, for fixed if the it is T^^ we show , . for that, can be no stronger, so T^, , . no stronger than thus I^(xn,a) look along go to zero some t]' (which T^, then x„ -* then for every a 4/M < .= max{|r|,77(r)} functions), equivalent to show > the • || T]i{r) fixed a, /^j(x„,a) does not the x*{q")-Xri{q") a contradiction. is Proof of proposition T/(l) < n, x'{q)-Xn{q) for sufficiently large we lemma for some For this use Suppose that some subsequence where . /^(x„,a) can be written as of three terms: V{^n{t)/a)l{i^„lt)\/a>i]dt Since each term so that the first is + positive, each term is less than T](xr,{t)/a)l[i^^(t)\/a<i]dt must go 1/M . 33 If to zero in the second n . + 7y(x„(l)/a). Pick n large enough is to go to zero, since t/ . positive except at the is strictly that the And argument Lesbesgue measure of the set e {t we zero, [0, 1] can find n large enough so > 1 : > 1/M} < 1/M x„(i)/a go to zero, it is necessary that, for n large enough, in(l)/a < 1/M. Picking n large enough so that all this happen at once, when we expand /^j(x„,a) in analogous fashion, we will find that (i) the first term in the sum is less than 1/M, since it is identical with the first term above, (ii) the for the third term to 1/M directly from our selection of n and (iii) the second term is less than 2/M, when we overestimate r]i{xn{t)/a) in the second term with the function that is 1/M if the argument is less than 1/A/ and 1 is the argument is between 1/M and 1 Since this is so for all n above some level, we contradict the supposition. term third than less is , , . Next we must show that, and Skorohod topologies. for show —+ suffices to if Xn -+ The that a:„ if any norm is no stronger than the is Since total variation in total variation, same in total variation, then the total variation T^ tj , is norm a (linear) is then !„ total variation — T^ > any true for a:„/a for topology, it Apply A2 - . scalar a > . sup norm, so the same must be true in convergence gives the result. As for the Skorohod page 123, problem 2] states the result for the L^ norm larger than the And dominated the sup norm. topology, Billingsley [1968, T topology, which immediately gives the result for For general . note that 77 , if — X in the Skorohod topology, then for n sufficiently large, sup^ kn(OI ^ supj \x{t)\ + 1/n Fixing any a , then, and evaluating /,,(xn — i, a) , we know that x„ > . the arguments of Convergence of will 77 J,,(x„ - Ji(xn — 77 , since means /i 5. can overestimate In A'+ , the Accordingly, the topology 77(r) sequence is sequential, is ^ if 2/(an). Ii(xn — 'q*{t) x,a) = t . described by the Prohorov and convergent in one of the two topologies + Convergence of and we are done. with rr]{K/a) weak topology is if the integral using the function x,a) to zero follows from the problem in Billingsley, Proof of Proposition metric. we \x{t)\/a x,a) to zero, then, follows immediately converges to zero, where in place of K/a = 2supj be bounded above by it suffices to and only show that a convergent if it is > 0, and, applying lemma A2, pick N sufficiently large so that /,,(x„ x,e) < e for n > N We claim that this implies that p(x„,x) < e For if not, then there exists some t such that ^nit + e) + e < x{t) or such that x„(< - e) - e > x{t). Suppose that the former is true, and suppose further that t e [0,1 - e]. See figure Al. In calculating Ir,{xn - 2:, e) when we integrate T]{{xn{t) - ^it))^) over [t, t + e], the integrand Suppose in the other. first that x^ x in T^ . Fix e - . . , exceeds 77(1) = 1. (Recall that e, a contradiction. If t > I - 77(1) e, = 1 is then x(l) 34 Hence the - x„(l) > e, and wlog.) integral the last is at least (summed) , term in some i /,,(xn Xn{t , - x, e) will exceed 1 again a contradiction. The - e) - e > x{t) is handled similarly. case where, for , ^ x in the weak /^(x„ -x,a) ^ 0, so that x„ > lemma A2. Conversely, suppose that Xn topology. Fix a constant a We -> x If will show p(x„,x) that — 0, then ^ it immediate is term in 7;,(x„,a) converges with the first (integral) large so that for n nondecreasing, we that to zero in n . - |xn(l) We , X(l)| -> 0, m, so that the . final N{m) be sufficiently that both x^ and x are let < 1/m. RecaUing p(xn,x) follows by need, therefore, only be concerned term. For a given integer > N{m) T^ can bound above by 1 m ^ a \ ^ Jt=i Now since x{t /max{x{k/m),Xn{k/m)] -mm{x({k -l)/'m),Xn{{k - l)/m)} -^ p(xn,x) — 1/m) — 1/m, < 1/m, max{x(<),x„(i)} < x(i+l/m)+l/m and so the summation given _^^ ( k=\ V^ m ^ From the shape of is bounded above by + l)/m) - x{{k is x{{k min{x(i),Xn(i)} bounded above by - 2)/m) + 2/m a (convex, equal to zero at zero, and increasing), this in turn 77 x{{k + l)/m) - xi{k - 2)/m) + 2/m \k=\ ^ 1 m For fixed a , the argument of r/ is /3x(l) + 2 a \ fixed, so this upper bound vanishes with m proving our assertion. Proof of the technical assertion in the proof of proposition x„ and T/n are as in the (2.3), then x„ techniques employed just above 1 /3x„(l) n \ a show — 1/„ — > that an + 2\ T^ 4. . We Fixing a upper bound on /^x„(l)-y„(l) J 35 show scalar a > need to /,,(xn — that if 0, the yn,a) is > , The second term vanishes for any a since a;„(l) — y„(l) < p{xn,yn) < \/n And . By assumption, a:n(l)//i(n) — 0, and thus (3x„(l) + 2)/(a/x(n)) — 0. Hence for any e there is A''^ such that for all n > N( to both sides ( (3x„(l)+2)/a < efi(n) Apply is monotone), and use the shape the first term vanishes for r-^y a: > > . of r?, for all for e < 1 to 7/ r; conclude that r7((3i„(l) n > N(. 36 + 2)/a) < T]{efi{n)) < er]{fi{n)) = en, A neighborhood of the function x in the Prohorov Another function y is within epsilon of the function x y is connpleteiy within the "sleeve" around x that is drawn. Note well the hatched part of the sleeve near the origin, which exists because we implicitly think of x as being extended for negative arguments having the value zero. Figure 1 : metric. if t+e t Figure A1 each . The dashed box of length epsilon. is a square whose sides are 7131 U53 Date Due OEC 8 A99A' ffi6.0 Lib-26-67 3 TDSD DD5 E22 Qfi5