Information and Sufficiency on the Stock Market Peter Harremoës Copenhagen Business College Copenhagen Denmark Email: [email protected] Abstract—Itiswell-knownthatthereareanumberofrelations leads to Bregman divergences and sufficiency lead to local 6 betweentheoreticalfinancetheoryandinformationtheory.Some proper scoring rules. The basic result is that any strictly local 1 of these relations are exact and some are approximate. In this proper scoring rule is proportional to logaritmic score. The paper we will explore some of these relations and determine 0 linkbetweeninformationtheoryandstatisticsisnowverywell under which conditions the relations are exact. It turns out 2 that portfolio theory always leads to Bregman divergences. established [6]. n The Bregman divergence is only proportional to information Convex optimization also appear in thermodynamics and a divergence in situations that are essentially equal to the type statistical mechanics where the goal is to extract as much J of gambling studied by Kelly. This can be related an abstract energy as possible from some physical system. The notion 7 sufficiency condition. of entropy obviously play an important role in both theories, 2 I. INTRODUCTION butthebestinterpretationhasbeendebatedeversinceShannon ] decidedtocallhisquantityentropy.Sinceallthesetheoriesare T The relation between gambling and information theory has related we also get a link between finance theory and physics .I been known since Kelly [1]. Later Kelly’s theory has been sothereisawholetopiccalledeconophysicswhereideasfrom s extendedtotradingofassets,butthelinktoinformationtheory physics are applied to economic systems. We hope that the c [ is weaker than in the case of gambling [2]. In both gambling presentpaperwillhelptounderstandtowhatextendquantities theory and more general portfolio theory logarithmic terms in finance are really proportional to quantities in information 1 v appear because we are interested in the exponential growth theory, statistics, or physics. 3 rate. In this paper we shall demonstrate that portfolio theory The general idea of using Bregman divergencesfor convex 9 consist of two parts. The general part is related to Bregman optimization was presented in [7]. In the present paper we 5 divergences and this part is shared with a number of other will develop the theory further. Therefore there will be some 7 convex optimization problems. If a sufficiency condition is overlap between then the presentation in [7] and the present 0 . imposedonthegeneraltheorywearriveata theorywherethe paper.Thesecondgoalofthispaperisapplythegeneraltheory 1 Bregman divergence reduces to information divergence. The to portfolio theory. 0 sufficiency is essentially equal to Kelly’s theory of gambling. 6 II. OPTIMIZATION The general theory of convex optimization and Bregman 1 : divergences has a number of important applications. In each Assume thatour knowledgeofa system can be represented v of the applications we get a strong link to information theory by an element in a convex set S that we will call the state i X if a sufficiency condition is imposed. Therefore sufficiency space. The simplest case of a state space is the simplex of r conditions will lead to strong relations between the different probability measures on a set. In quantum information theory a applications. thestatespaceisthesetofdensitymatricesonaHilbertspace. In information theory an important goal is to compress. For states s0 and s1 and t ∈ [0,1] the convex combination As long as we restrict to uniquely decodable codes we get a (1−t)·s0+t·s1 is identified with the mixed state where s0 Bregmandivergence.Thesufficiencyconditioncorrespondsto is taken with probability 1−t and the state s1 is taken with allowingcodewordsrealvaluedlengthwhichisrelevantwhen probability t. The pure states are the extreme points of the weallowblockcodeswithnoupperlimitontheblocklength. state space.For simplicity we will assume thatthe state space Thisleadstothewidespreaduseofinformationdivergencein is a finite dimensional convex compact set. information theory. The link between information divergence Let A denote a subset of the feasible measurements such and the notion of sufficency was emphazied already by Kull- thata∈A mapsS into a distributionon the real numbersi.e. back and Leibler in 1951 in the paper entitled “Information a random variable. The elements of A may represent actions and Sufficiency” [3]. (decisions) that lead to a payoff like the score of a statistical In statistics the idea of scoring rules has its roots in the decision,theenergyextractedbyacertaininteractionwiththe 1920’s in the Dutch book theorem by Ramsay and de Finetti. system, (minus)the lengthof a codewordofthe nextencoded McCathy [4] studied scoring rules in a more systematic way inputletterusingaspecificcodebook,ortherevenueofusing and Dawid, Lauritzen and Parry [5] have recently extended acertainportfolio.Iftheactionaisappliedtothestate sthen the notion of proper local scoring rules. Proper scoring rules we get a random variable a(s) that we will allow to take values in R ∪ {−∞}. For each s ∈ S we define F (s) = Theorem 4. If (t1,t2,...,tn) is a probability vector on the supa∈AE[a(s)]. Without loss of generality we may assume statess1,s2,...,sn with s¯= ti·si and aopt is the optimal thatthe setof actionsA isclosed so thatwe mayassume that action for s¯then P thereexistsa∈AsuchthatF (s)=E[a(s)] andin thiscase C ≥inf t ·D (s ,a)+D (s¯,a ). we say that a is optimal for s. We note that F is convex but F a i F i F opt F need not be strictly convex. If a is an action anXd s is optimal then opt Definition 1. If F (s) is finite the regret of the action a is supD (s ,a)≥C +D (s ,a). F i F F opt defined by i DF (s,a)=F (s)−E[a(s)] (1) III. SUFFICIENCY Proposition 2. The regret DF has the following properties: Let(sθ)θ denoteafamilyofstatesandletΦdenoteanaffine transformation S → T where S and T denote state spaces. • DF (s,a)≥0 with equality if a is optimal for s. ThenΦissaidtobesufficientfor(s ) ifthereexistsanaffine • If aˆ is optimal for the state sˆ = ti · si where transformation Ψ:T →S such thaθt Ψθ (Φ(s ))=s . θ θ (t1,t2,...,tℓ) is a probability vector thPen We define a transformation Φ to be an isomixture if Φ ti·DF (si,a)= ti·DF (si,aˆ)+DF (sˆ,a). has the form Φ = ki=1pi · Φi where (p1,p2,··· ,pk) is a probability vector and Φ is a isometry, i.e. a bijective i • Xti·DF (si,a) is minXimal if ais optimal for ti·si. transformation of thePstate into itself. We say that the regret IfthePstateiss1 butoneactsasifthestateweres2 onPesuffers DF on the state space S satisfies the iso-sufficiency property a regret that equals the difference between what one achieves if and what could have been achieved. DF (Φ(s1),Φ(s2))=DF (s1,s2) (5) Definition 3. If F (s1) is finite the regret is defined by for any isomixture S → S that is sufficient for (s1,s2). The notion of sufficiency as a property of divergences was DF (s1,s2)=inafDF (s,an) (2) introduced in [9]. The crucial idea of restricting the attention wheretheinfimumistakenoveractionsa thatareoptimalfor totransformationsofthestate spaceintoitself wasintroduced s2. in [10]. The center of a convex set S is the set of point in S that If there exists a uniqueaction a such that F (s)=E[a(s)] are invariant under isometries of S. Note that the center is then F is differentiable which implies that the regret can be convexand non-empty[11]. If the center of the state space is written as a Bregman divergence in the following form not a point there are many Bregman divergences that satisfy DF (s1,s2)=F (s1)−(F (s2)+hs1−s2,∇F (s2)i). (3) the sufficiency condition. In the context of forecasting and statistical scoring rules the Proposition 5. Let G denote the set of isometries of a state use of Bregman divergences dates back to [8]. spaceS andletµ denotethe Haarprobabilitymeasureon G. Bregman divergences satisfy the Bregman identity Let Φ denote the projection s → g(s) dµ. Let F denote a concavefunctiononthecenterofS.ThenDF (Φ(s1),Φ(s2)) ti·DF (si,s˜)= ti·DF (si,sˆ)+DF (sˆ,s˜) defines a Bregman divergence onR S that satisfies the iso- but ifXF is not differentiabXle this identity can be violated. If sufficiency condition. the state s2 has the unique optimal action a2 then Proposition 6. Assume that S is a state space. If the diver- genceD satisfiestheiso-sufficiencypropertythenthereexists F(s1)=DF (s1,s2)+E[a2(s1)] (4) a F˜ suchFthat sothefunctionF canbereconstructedfromD exceptforan F affinefunctionofs1. SimilarlythedivergenceDF isuniquely DF˜(s1,s2)=DF (s1,s2) determined by the function F. and F˜(Φ(s))=F˜(s) . Consider the case where the state is not know exactly but If the state space is a one dimensional simplex then the weknowthats∈S forsomesetofstates.Theminimaxregret only sufficient transformation is the reflection and the above of the set S is defined as conditiononF issufficienttoconcludethatEquation5holds. C =infsupD (s ,a). F F i a i Proposition 7. If the state space has the shape of a ball then Using general minimax results we get anyfunctionF ontheballthatisconcaveandinvariantunder rotations satisfies the iso-sufficiency condition. C =supinf t ·D (s ,a) F i F i Proof: Assume that the isomixture Φ is sufficient for ~t a Xi {s0,s1}. Then Φ is also sufficient for any affine conbination where the supremum is taken over all probability vectors ~t ofs0 ands1. Inparticularwe mayreplaces0 ands1 byaffine supported on S. This result can improved. combinationsfor the form sti =(1−ti)·s0+ti·s1 that are extreme points in S. Since Φ is assumed to be sufficient it Example 11. A specialasset is the safe asset wherethe price maps s into an extreme points. Hence Φ acts as a rotation relative is 1 for any feasible price relative vector. Investingin ti on the intersection of the state space and the affine span this asset correspondsto place the moneyat a safe place with of s1,s2 and U. Since F is invariant under rotations the interest rate equal to 0 % . divergence D is also invariant under rotations implying that F A portfolio is an asset given by a probability vector~b = DFTh(Φes(ism1)pl,eΦst(csa2s)e)=oDfaFb(asl1l,iss2a)n.interval,whichcorresponds (b1,b2,...,bk) where for instance b5 =0.3 means that 30 % ofthemoneyisinvestedinassetno.5.Thetotalpricerelative to the probabilitymeasureson a binaryalphabet. This special case was discussed in [10]. The balls in dimensions 2, 3, and isX1·b1+X2·b2+···+Xk·bk = X~,~b .Ifanassethasthe property that the price relative is only positive for one of the 5 correspond to density matrices of a 2 dimensional Hilbert D E feasible price relative vectors,then we may call it a gambling space over the real numbers, over the complex numbers, and asset. For any set of possible assets we may extend the set over the quarternions. of assets by a numberof idealgambling assets so that any of We say that the states s0 and s1 are orthogonal and write the possible assets can be written as a portfolio of the ideal s1⊥s2 if there exists an affine function φ : S → [0,1] such gambling assets. This can be done without changing the set φ(s0) = 0 and φ(s1) = 1. The following theorem can be of feasible pricerelative vectors.Thereforethe set of possible proved by the same technique as [7, Thm. 4] except that we portfolios may be considered as a convex subset of a set of will make sufficientprojectionsby takingthe mean actionsof portfolios of some ideal gambling assets. a groups equipped with the Haar probability measure. We now consider a situation where the assets are traded Theorem 8. Assume that the state space S satisfies the once every day. For a sequence of price relative vectors following properties: X~1,X~2,...X~n and a constant re-balancing portfolio ~b the 1.Forandtwopurestatess1ands2thereexistsanisometry wealth after n days is of S such that Φ(s1)=s2. n 2. For any three pure states s1,s2, and s3such that s1⊥s3 Sn = X~i,~b (6) ands2⊥s3 thereexistsanisometryofS suchthatΦ(s1)=s2 iY=1D E and Φ(s3)=s3. n 3.Thestatespacehasatleastthreeorthogonalpurestates. = exp log X~i,~b (7) 4. Anystate can be written as amixture oforthogonalpure Xi=1 (cid:16)D E(cid:17)! states. = exp n·E log X~,~b (8) If the regret D satisfies the iso-sufficiency property given F where the expectation is t(cid:16)aken whith rDespectEtio(cid:17)the empirical by Equation 5, then D is uniquely determined except for a F multiplicative factor. distributionofthe pricerelativevectors.Here E log X~,~b Remark 9. Condition 4 seems to be redundant, but we have is proportionalto the doubling rate and is denothed WD ~b,PEi not been able to prove this. where P indicates the probability distribution of X~. Ou(cid:16)r goa(cid:17)l When the state space is a simplex the uniquely determined istomaximizeW ~b,P bychoosinganappropriateportfolio divergenceisinformationdivergenceandwhenthestatespace ~b. In [2] and [7] it(cid:16)was(cid:17)tacitly assumed that a unique optimal isdensitymatricesonacomplexHilbertspacewegetquantum portfolio exists but this is not always the case. Here we will relative entropy. not assume uniqqueness. Lemma 10. Assmue that the state space satisfies the condi- Definition 12. Let~b1 and~b2 denote two portfolios. We say tionsin Theorem8. Ifs0⊥s1 thenanyoptimalactiona fors1 that ~b1 dominates ~b2 if X~j,~b1 ≥ X~j,~b2 for any j = satisfies E[a(s0)]=−∞. 1,2,...,n.Wesaythat~b1DstrictlyEdomiDnates~b2Eif X~j,~b1 > Proof:Since s0 ands1 are orthogonalandthe conditions in the previoustheoremis fulfilled the we have thatthe regret X~j,~b2 for any j =1,2,...,n. D E prersotpriocrtteidontaoltthoeinlifnoermseagtimonendtiv{etrg∈en[0c,e1,]b|u(t1in−fotr)msa0ti+ontsd1iv}eirs- D ForaEvector~v =(v1,v2,...,vk)∈Rk thesupportsupp(~v) is the set of indices i such that v > 0. We note that m∞geun.mcHeieesnqctuaeakleisnn∞faov(feForr(aoscr0tt)iho−ongsEotnh[aaalt(dasir0set)r]io)bpu=ttiimo∞nasl fswoorhDesr1Fe.(Ttshh0ee,rsien1f)foo=rie- iif∈~b1sustprpictl~by2 dosmucinhattehsat~b~b21isftraicntdlyoidnolymiinfattehser~eei ewxhisetrse a~eni E[a(s0)]=−∞ for any action a that is optimal for s1. denotesthe(cid:16)i’t(cid:17)hbasis vector.The consequenceis that we may removeassetsnumberiif~e isstrictlydominatedbecauseone i IV. PORTFOLIO THEORY willneverputanymoneyonthatparticularasset.Similarly,~b1 Let X1,X2,...,Xk denote price relatives for a list of k dominates~b2 ifandonlyifthereexistsani∈supp ~b2 such assets.ForinstanceX5 =1.04meansthatassetno.5increases that~b1dominates~ei.Wedonotdecreasethemaxima(cid:16)ldo(cid:17)ubling its value by 4 %. rate by removing assets that are dominated, but sometimes assets that are dominated but not strictly dominated may lead measuresare projected (reverse informationprojection)into a to non-uniquenessof the optimal portfolio. convex set. Here we should note that information divergence is convex but not strictly convex in the second argument. Definition 13. A set A of assets is said to dominate the set Therefore the reversed information may be non-unique. of assets B if anyasset in B is dominatedby a by a portfolio of assets in A. V. SUFFICIENT PORTFOLIOS Proposition 14. If~b0 is optimal for the distribution δ~v then Lemma 16. Assume that there are only two price relative the support of~b is a subset of the support of ~v. vectorsandthatthesetofassetsisminimaldominating.Ifthe Proof:IfP =δ thenE log X~,~b =log ~v,~b .The Bregman divergence ~v portfolio~b isaprobabilitydisthributDionovEeirstocksDsoifEwelet W ~b ,P −W ~b ,P (11) P Q ~b∗ denote the conditional distribution of~b on the support of (cid:16) (cid:17) (cid:16) (cid:17) ~v. Then isproportionaltoinformationdivergenceD(PkQ)thenthere log ~v,~b0 ≤log ~v,~b∗ are only two gambling assets. D E D E with equality if and only if the supportof~b is a subset of the Proof: Let sisupapsourbtsoeft~vo.f Tthheerseufpoproer~bt o=f ~~vb.0 implies that the support of~b X~ = (X1,X2,...,Xk) Let~b denote a portfolio that is optimal for P. The regret Y~ = (Y1, Y2, ..., Yk) P of choosing a portfolio according to Q when the distribution denote the two price relative vectors. If P = (s,t) then the is P is given by the Bregman divergence vector~b=(b1,b2,...,bn) is log-optimal if and only if W ~b ,P −W ~b ,P . P Q X Y i i (cid:16) (cid:17) (cid:16) (cid:17) s + t ≤ 1 If~bQ is not uniquelydetermined we take a minimum over all b1X1+···+bkXk b1Y1+···+bkYk ~b that are optimal for Q. for all i ∈ {1,2,...,k} with equality if b > 0. Since we i have assumed that none of the assets are dominated by other Example 15. If the assets are orthogonalgamblingassets we portfoliosonlytwooftheseinequalitiescanholdwithequality. get the type of gambling described by Kelly. There will be one-to-onecorrespondencebetween price relative vectors and Therefore we may assume that only b1 and b2 are positive. Hence we may assume that there are only two assets. assets. For a probability disttribution P over price relative vectors the optimal portfolio ~b is a vector with the same Let δ1 denote the measure concentrated on X~ and let δ2 coordinates as the probability vePctor P. We have denotethe measureconcentratedonY~. Since the measuresδ1 W ~bP,P −W ~bQ,P =D(PkQ) (9) a−n∞dδ.2Naorwe orthogonalLemma 10we have thatW (cid:16)~bδj,δi(cid:17)= (cid:16) (cid:17) (cid:16) (cid:17) so the sufficiency condition is fulfilled in gambling. W ~b ,δ = E log X~,~b δj i δi δj If a set of possible assets it embedded as a subset C in a (cid:16) (cid:17) = log hX~ ,D~b Ei set of ideal gambling assets then C may be identified with i δj a convex set of probability distributions. Now maximizing D E W ~b,P overpossibleportfolios~bisthesameasminimizing so that X~i,~bδj = 0. Since the support of ~bδi is a subset the(cid:16)regre(cid:17)tgivenby(9)overQ∈C inthesetofportfoliosover of the suDpport ofEX~i we have that~bδi⊥~bδj. Therefore~bδ1 and ideal gambling assets. Therefore~b may be identified with a ~b must be proportional to the basis vectors. Since~b and Q δ2 δ1 reversed information projection of Q on C. ~b are vectorsin a 2-dimensionalspace and their coordinates δ2 As proved in [2] the regret satisfies arenon-negativewe havethat~b mustproportionalto a basis δi vector. Since X~ ,~b = 0 for i 6= j we have that X~ is W ~b ,P −W ~b ,P ≤D(PkQ). (10) i δj i P Q parallel with~bD. E (cid:16) (cid:17) (cid:16) (cid:17) δi In the set of portfolios over ideal assets there is a on- Theorem 17. Assume that none of the assets are dominated to-one correspondence between mixed states and portfolios. Therefore maximizing W ~b,P over~b in the original set of by a portfolio of the other assets. If the Bregman divergence portfolioscorrespondsto m(cid:16)inim(cid:17)izingthe regretW ~b ,P − W ~b ,P −W ~b ,P (12) P P Q W ~b ,P over Q which again corresponds to m(cid:16)inimiz(cid:17)ing (cid:16) (cid:17) (cid:16) (cid:17) Q is proportionalto information divergence D(PkQ) the mea- D((cid:16)PkQ)u(cid:17)ndertheconditionthat~b ∈C inasetofportfolios sures P and Q are supported by k distinct price rela- Q on orthogonal gambling assets. The inequality (10) therefore tive vectors of the form (o1,0,0,...0), (0,o2,0,...0), until states that information divergencedecreases when probability (0,0,...o ). k Proof: Assume that there exists a constant c > 0 such VI. CONCLUSION that The link between portfolio theory and information theory W ~b ,P −W ~b ,P =c·D(PkQ). (13) works on two levels. Parts of the theory can be stated and P Q proved on the level of convex optimization, where Bregman (cid:16) (cid:17) (cid:16) (cid:17) If~b =~b then divergences and related concepts play a central role. If we P Q further impose a sufficiency condition we have, essentially, W ~b ,P −W ~b ,P =0 to restrict our attention to gambling as described by Kelly. P Q Adding certain assets that are dominated does not make any (cid:16) (cid:17) (cid:16) (cid:17) significant changes to the theory. In the case of gambling and D(PkQ)=0 and P =Q. Therefore the mapping P → ~b isinjective.Thevectors~b formasimplexwithk extreme the correspondence between portfolio theory and information P P theory becomes perfect. Therefore the link between general points.ThereforethesimplexofprobabilitymeasuresP hasat portfolio theory and information theory is convayed by gam- mostk extremepoints,soP issupportedonatmostk distinct vectors that we will denote X~1,X~2,...,X~k. bling theory. Assume that X~ and Y~ are two vectors of price relatives. Information divergence was introduced by Kullback and Leiblerinthepaperentitled“OnInformationandSufficiency”. Then Equation 13 holds for probability vectors restricted to In the present paper we have made the notion of sufficiency the set X~,Y~ . From Lemma 16 it follows that X~ and Y~ more explicit for portfolio theory. The introduction of ideal are orthogonal. Therefore all the price relative vectors are n o gambling assets paralellels the use of microscopic states as orthogonal,andhavedisjointsupports.Sincethepricerelative opposed to macroscopic states in physics. For microscopic vectorshavedisjointsupport,anassetcanonlyhaveapositive states we have reversibility and conservation of energy. Sim- price relative for one of the price relative vectors. Therefore ilarly, gambling corresponds to two-person zero sum games each price relative vector has one asset that dominates any where money is the conserved quantity. As we have seen other asset in the support of the price relative vector. Since these correspondencies are consequences of the sufficiency we have assumed noneof the assets are dominatedeach price condition. relative vector is supported on a single asset. If the price relative vectorsare as in Theorem 17 we are in ACKNOWLEDGEMENT the situation of gambling introduced by Kelly [1]. ThaauthorwanttothankPrasadSanthanamforinvitingme to Electical Engineering Department, University of Hawai’i, Corollary 18. Assume that the Bregman divergence where this paper was written. W ~bP,P −W ~bQ,P (14) REFERENCES (cid:16) (cid:17) (cid:16) (cid:17) [1] J. L. Kelly, “A new interpretation of information rate,” Bell System satisfies the sufficiency condition for probability measures P Technical Journal, vol.35,pp.917–926,1956. and Q supported on k ≥ 3 price relative vectors. Then the [2] T. Cover and J. A. Thomas, Elements of Information Theory. Wiley, 1991. setofpossibleassetscontaink gamblingassetsandanyother [3] S.KullbackandR.Leibler,“Oninformationandsufficiency,”Ann.Math. asset is dominated by a portfolio on the gambling assets. Statist., vol.22,pp.79–86,1951. [4] J.McCarthy, “Measures ofthevalueofinformation,” Proc.Nat.Acad. Example19. IftheBremandivergencesatisfiesthesufficiency Sci.,vol.42,pp.654–655, 1956. condition and one of the assets is the safe asset then there [5] A. P. Dawid, S. Lauritzen, and M. Perry, “Proper local scoring rules exists a portfolio~b such that b ·o ≥1 for all i. Equivalently ondiscretesamplespaces,”TheAnnalsofStatistics, vol.40,no.1,pp. i i 593–603,2012. bi ≥ o−i 1 which is possible if and only if o−i 1 ≤ 1. One [6] I. Csiszár and P. Shields, Information Theory and Statistics: A Tuto- say that the gamble is fair if o−1 = 1. If the gamble is rial, ser. Foundations and Trends in Communications and Information superfair,i.e. o−1 <1,thenthepiortfolioPb =o−1/ o−1 Theory. NowPublishers Inc.,2004. gives a price relatiive equal to P o−1 −1 >i 1 inidependeitly [7] tPh.eHTahrreemEoiëgsh,th“PrWooprekrshsocporinogn aInnfdormsuaftfiiocinencTyh,”eorineticProMceetehdoindgs ionf P i P Science and Engineering, ser. Series of Publications B, J. Rissanen, of what happens, which is a Dutch book. (cid:0)P (cid:1) P. Harremoës, S. Forchhammer, T. Roos, and P. Myllymäke, Eds., no. Report B-2015-1, University of Helsinki, Department Corollary 20. Assume that there are at least three distinct of Computer Science, 2015, pp. 19–22, an appendix with proofs price relative vectors. The Bregman divergence (14) satis- only exists in the arXiv version of the paper. [Online]. Available: fies the sufficiency doncition if and only if W ~b ,P − http://www.cs.helsinki.fi/u/ttonteri/pub/witmse2015proceedings.pdfhttp://arxiv.org/abs/1507.070 P [8] A. D. Hendrickson and R. J. Buehler, “Proper scores for probability W ~b ,P =0 implies P =Q. (cid:16) (cid:17) forecasters,” Ann.Math.Statist., vol.42,pp.1916–1921,1971. Q [9] P. Harremoës and N. Tishby, “The information bottleneck revisited or (cid:16) (cid:17) how to choose a good distortion measure,” in Proceedings ISIT 2007, Proof: If Equation 9 does not hold then we do not have Nice. IEEE Information Theory Society, June 2007, pp. 566–571. sufficiency so the set of possible portfolios can be identified [Online]. Available: www.harremoes.dk/Peter/flaske2.pdf with a convex and proper subset of the set of all portfolios [10] J.Jiao,T.C.amdAlbertNo,K.Venkat,andT.Weissman,“Information measures: the curious case of the binary alphabet,” Trans. Inform. on a set of gambling assets. Then we just have to find to Theory,vol.60,no.12,pp.7616–7626, Dec.2014. distributionsP andQthathavethesamereversedinformation [11] T.-C.Lim, “Thecenter ofaconvex set,” Proceedings of the American projection into the set of possible portfolios. Mathematical Society, vol.81,no.2,pp.345–346,Feb.1981.