Subscribe to DeepL Pro to translate larger documents. Visit www.DeepL.com/pro for more information. 175 S HR LES F 0 NC TI0 NS C0 N VEX ES ET LES I N É G A L I TÉS ENTRE R E LES VALEIJ RSg M0 YEN NES' R J. L. W. Y. JENSEN coPEmR*Oos. i . Functions as:ces and concares. DeJnition. Examples. In his famous Ana lyse algé bri qu e (note II, pp. 4ç 2-Çq) Cxc-car demonstrates that "the geometric mean between several numbers is always less than their algebraic mean". The method employed by Axncev is extremely elegant, and is ä pa 8é without change in all treatises on algebraic analysis. It consists, as is well known, in the fact t h a t , from the inequality where a and h are rositive numbers, we are led to the analogous inequality for four numbers, namely and the following, for 8, i 6, ... 2 numbers, after which this number, by an artifice, is reduced to an arbitrary lower number, n. This simple method was my starting point in the following investigations, which lead, by what is in fact a very simple and elementary route, to general and not unimportant results. ' Lecture delivered at the Danish Mathematical Society on January Acte mothemntieo. 30. Imprimé le t9 déeezisbre t90b. 9°I 176 J. L. W. ¥ . Jensen. I'll introduce the definition snivants. When a Î'onction p(z), real, finite and uniform, of the real variable z, satisfies in a certain inter- valle t h e inequality we say that p(z) is a connected function in this interval. If, on the other hand, Cf z) satisfies the inequality (-) P(-) + r(z) ñï*r("" !') we say that p(z) is a concave function. Furthermore, we assume that these inequalities do not reduce to t h e g i v e n interval, to equality (j) In this case, the function p(z) is said to be -linear- in the given interval. This expression has been adopted because 1 equality (3) is satisfied by It follows from these definitions that -1-linear functions form a transition from the class of convex functions to that of concave functions, and must be considered as limiting cases common t o both classes. From the definitions given, it immediately follows that qi:e - p(s) is concave, when p(z) is connected, and vice versa. In what follows, it would be auÎfisant to consider only convexB functions, since it' s so easy to switch from one clause t o another. As, however, it may be advantageous to consider both clauses of functions, con- cave functions will be mentioned from time to t i m e ; but it should be remembered, even when this is not 8always recalled, that every pro-position relating to convex functions corresponds to an analogous proposition for concave functions8. As t h e main purpose of this research is to present a series of rlalités d'un caractère général, comprenant comme cas particuliers presque toutes les inégalités' jusqu'ici connus, nous allons développer d'abord les théorèmes néeesanires propres a ce but, avant d ' entreprendre l'étude plus approfondie des fonctions convexes elles-mémes. On convex functions and inequalities between mean values. 17 7 It will first be necessary to state some propositions that immediately follow from the definitions, and to give some examples of the classes of functions defined. A sum of convex or -linear functions is connected when at least one of the functions is connected. If p(z) is connected, and c is a positive constant, ep (z) is connected. These propositions are also true for concave functions. Examples: I°. z' is a connected function in any interval. So a+ ôz -]- es' is connected or concave, depending on whether c is positive or negative. 2º. l - l is linear " in any interval that does not include the zero, but connected otherwise. So where c are positive constants, is connected or -linear in any interval, connected if the interval includes one or more dea values z, , z"......................, z" -linear-, if that doesn't happen. 3º. Inequality which occurs for o N b > o, and for any value of p greater than ä i , is If z and y are positive, and z > y, we deduce from this inequality (g) ")' - So z*, for positive z values and a p value greater than i , is a related function. and y to - and From the last inequality it follows, by changing z , 2sy pat a very elementary method. Acta mclhe'natiaa. 30. Imyritzt é te 20 üé ceæ6ze I 90o. 2ö 178 J. L. ,". À'. Jensen. So z *, when j' > i , is a related function }'for positive values. tives de z . If, on the other hand, we put ttans ("} z^ and j/" 'i lii rlace of z and y, we obtain z" is therefore a concave function, for positive z, when o < y < i . From (§) it follows which is still true for y' =- I. Thus .z°" is also O convex when In summary, for positive values of z , we see that s' is convex when y is than i , or negative, and concave when If y r'us large =- i , the function is "linear", a s we saw above. 4°. | Çn -y- /'z' | is convex ponrconcave for ab < o, as long as the binòme soua the radical sign remains positive. ç°. e' * is a convex function in any interval, while log z is concave in the interval (o , -}- en). 6º. Differential calculus provides us with a general way of deciding whether functions that can be differentiated are eonvex or concave. If /'(z) is real and uniform in a given interval, and has a finite and continuous second derivative, then, by virtue of Eori.x's theorem, we have 2 .r' being an intermediate number between x and y. So / (z) will be connected as long as Ç"(z) is positive, concave as long as c(ue {"(z i is nepatiVe. The geometrical significance of this fact is obvious. lii e8et, if a conveze function p(z i is siisceptihÎe r1'uno representation ¿-eoinétrique en coor- rlonnees rectangulaires l'équation y - ÿ t z ) cléfinit ii ne courbe, turning its convexity towards negative y. On convex functions and inequalities entFe mean values. Remargue. 179 The following three inequalities (Z) -|- /(ÿ) > 2 ( ÿ 'zy), '/(z)'/(y) > j/(ÿzy)j\ s and y positive; z eî y positive, (z) positil; can be reduced to (i) by simple substitutions, namely, respectivelyment: log/(s)' P(*)' /(e") - ( s), dog (e") p(s). z. Generalization o f inequality (i). Suppose that p(z) is any connected function in a given interval, and that z, , z, , z, , . . . are all located in this interval or at its limits. From (i) it follows that P(-.) + P(-.) + (-.) + P(-.) * 2P(*' *') + 2P(°' t * ) +n 4 and generally, m being a positive integer u=1 as can easily be seen from t h e complete indnction.If then ri is a positive integer, and if we choose m such that 2' ri, we can pose Then we find 180 J. L. W. V. Jenaen. which is a generalization o f inequality (i), used to define 1 convex functions. Clearly, a similar inequality applies to con- cave functions: all you have to do is reverse the sign o f inequality. For -li- neara- functions, the inequality simply becomes an equality. Snpposon8 now that p(z) is an eottfinite and connected function in a certain interval. We know that such functions exist from the previous examples. Let ri u , -{ri, -[- ... -{- xq, where all nq are positive integers. It follows from (4), by choosing the z's i n a suitable way, Let and , o, , . . . eq be any positive numbers whose sum is a, and let 'i grow indefinitely, but in such a way that lim n' O' lim -' - " HO ' . ' lim -^' - "*' taD ' which will result in lim and by anite, p(z) being continuous by assumption, which demonstrates the following theorem: t h e inequality oîi x , x , ... repi eseitfenf numbers all within the interval, and yes a, , a, , . soul positive numbers, but any numbers. On oonrex functions and inequalities between average values. 181 For concave functions, the sign of the inequality must be reversed. This proposition is so general that perhaps a l l known inequalities between mean values are included as cases. very special. V O ppliCotions of the formula ($). 3 In the following, the numbers represented by a, h, c, . . . e , $, T, ... will be assumed positive. In formula (ç), let's say p(z) z^, p > i , so p(z) is connected, as we' ve seen. Then we have or (6) where all z are positive. Inequality (6) reduces to a simple identity for z, z , =... z . It also holds for whereas it must be reversed for For e 1 i we find a result previously given by Mr. R. SIHON.' By making p -- 2, and replacing a, by a*, z, by , we find (') formula by CxUCTiz {loc. ciL, p. 4 Ç), who gives a very different de- monstration. Posing in (/) e} z} for o, , a, z ' for ô " we have ' Über einige Ongleichungen, Ze its cli ri f t f ü r ôf ath. u. P h ys i - 33 P 37. -8 - 182 J. L. W. V . Jensen. d'oil follows that the harmonic mean between several positive numbers is smaller than their algebraic mean. It is easy to generalize formula i7) Replace in (6) s:,by andextract the root p"''° from the deua members, we will have which can be written in symmetrical form z, , r, being positive constants with sum I . Raising the two members of t h e above inequality to the power z' "°'e where z' is positive and < I , and multiplying byr , we find which demonstrates t h e following inequality zi , z, , x, being positive constants with sum i. this way, we demonstrate b y complete induction Continuing in (8) where z are positive constants8 with sum i . v In formula ( y), let's put a, h* instead of a, , and a* b* instead of in place of fi, where z and y are any real numbers, and for . . . simplify writing, S(x) - N o, ô( . We have: s('j'))"<s,)s(,) On convex functions and inequalities between mean values. and log N(z) is therefore a function The fundamental relation (ç) then gives: 183 of z in the interval (- en , -J- c'o). or S (9) 2, we have u v o V * For w *(s) (s(s,))-"(*(si')"' " or (* -))"'-*:< (*(-.')"-*(s(-,))"-"Hence OR POCOIO 1 From (I o) it follows that the function increases, z , remaining invariant. ( /) *' is never decreasing when z If in (i o') we permute z , and z" 184 J. L. W. V. Jensen. we must assume z, N z > z, , and conclude that $p(g ') ' " never increases as z decreases. This demonstrates the following proposition: The numbers a , , o , , . .. bi , by , . .. being positive, z being any real variable, and z any real constant, the fonetion is monotonic and never decreases as z increases, either in the interval ( c 'o , z ), or in the interval (z, , -]- c'o). We have &(z, - e) c 'f(z, -{- e), c being positive. The last part of the proposal results from the fact that as mentioned above. This proposal includes, as special cases, a number of tions of SciiröàIILCH ': mustri positive numbers ' x , S, y ... 2, and S$a- f- Q -\- . . . + I-, we'll have V and Moreover, BisNAYxE ' has stated, without demonstration, a proposition that is very close to our above proposal. Later Mr. R. SIäfON has published a demonstration o f a special case. ' Über Mitlelgrössen verschiedener Œdnungen, Z ei ta e h r i f t fü r Cf ath em a ti k u n d Physik, 3. P 3OI, i 8J 8. ' Société philomatique de Paris, Extracts from the minutes of meetings for 1 year 4 p. 6/, Paris 4 ' At the 'quoted' location. On convex functions and entFe lea valenra mean inequalities. 185 SCHLÖMILCH also determines limit values 1 For our function, it is easy to find that i and where b is the largest of the numbers b, , ô , , . .. , ô . To apply another application of the fundamental formula (3), let ç (z) log s, p(z) is concave and we then have II) which is a generalized form, due to Mr. L. J. ROGERS', of the classical proposition about the geometric mean. Another inequality of a similar nature can be found by noting that z log z is connected for positive z. Here it is (12) The special case where o, =- o, ... =- a, i was also given by Mr. Roczns. These examples should suffice to show just how fruitful formula (ç) can be. Acta trt4tfiei'irtticn. 30. Imprim ë le 20 décembre 90c. 186 J. L. W. V. Jensen. Clearly, in the 1e8 preceding formulas, when ja, Za,z" etc., are convergent series, we can increase ii indefinitely, thus obtaining a n inequality between certain infinite series. We can use the formula ( i i ) in series t h e o r y in a different way. let fib , Xb z . ... Zb" be contiguous series with positive terms, and let al , al , . . . a positive constants whose sum is i , the series will also be com urgent. The proof of this proposition can be deduced immediately from (i i) by making the b's dependent o n a new index v, posing a, -{- o, -}-i , and summing in both members for v i , c , .... ltOOtÎOîtS SttY Îe calcul intégral. There are still other cases where ri 4 can be made to grow indefinitely. Remembering the definition o f an integral defined as the limit of the values o f a sum, the above gives us a whole range o f interesting inequalities between integrals. Suppoaona that a(z) and f(:r) area functions integrablea in the inter-valle (o , i), and that a(z) is constantly positive. Inequality (ç) gives where p(z) is assumed to be continuous and connected in the interval (p , , p,), p, and g being the lower and upper limits of the function /(z) in the interval (o , i). Now, we know that the product of two integrable functions is an integrable function, and that an integrable function, by substitution in a continuous function, gives an integrable function. So, if we allow zt to grow indefinitely On eonvex functions and inequalltys between mean values. It goes without saying that the J 187 in this formula by corresponding integrals f. By analogy with the previous formulas, examples o f the application of formula (ç') fa(z)dz) f a (z) b(z)dz (o(s))*ôz Iwill cite the following f a(z)(((z))-dz, (à(s))*rfs, /(s) positive, @ I; b(s) intëgzable and positive; lea a(z) being positive and integrable, and x, -{- x, -{- . . -]- x, - I ; fr c) log è{s)d= O(*)1Og^t=@* Of these formulas, the second and fourth are generalizations of wellknown formulas. By positing a(z) i in the last one, we find a ScnLöMILca formula that is particularly interesting because it leads to an important result in the search for the zéPOB of a Txrron series, as I'll show elsewhere. 3. Further study of conne:ces functions. After demonstrating the usefulness of the notion of "connected function", we return to t h e study of such functions in general. In §z, we showed that formula (4) has li£iu p£lur any connected function 188 If we make n J. L. W. \ '. Jensen. this, we find Putting - d in place of d, and assuming z -]- ad and .r - nd within t h e given interval, we find As a result of the definition, the inequalities ('x) and (§) can be written as ||V |V | Assuming then that p(z) has a finite upper limit p in the given inter-valle, we have, taking m i from which it follows, if we decrease o to o, that n grows indefinitely, but slowly enough so that z + u'i does not leave the interval, km (p(z -}- 'i) - p(z)) o. This demonstrates the following proposition: On convex functions and inequalities between mean values. 189 Formula (Ç) therefore applies to any similar function. We could deduce the following, but it's }a1simpler to start from formula (y). If we put inthis iithe pl:ice of à, we find, now assuming o /'osi/t/ which becomes, by makingconverge to any positive number, this, smaller | V | than i This formula shows that "+never grows when d decreases, and that this quofient remains constantly greater than o' is positive, but i n any case arbitrary. 0 - ') or It follows that ' This proposition also applies to a concave function, by replacing it with >upper limit" by s lower limit". From the fact that a - linear function s can be considered as a special case of two clas8ea of functions, it follows: A "linear function" that has either an upper or a lower limit in a given interval is continuous. From this result we can easily conclude the following proposition: a "linear function" having either an upper or a lower limit in a given interval always has the form a -}- bs dttnB this interval, 'i and b being constants. This fully justifies the name we have introduced. 190 exists, and we c a n also see that is also determined. This theorem is thus demonstrated: A J'onctimi con "er 'y(x), qiti't in a certain interval a finite upper limit, has a derivative function tattt ii right, f' (x), than gaucH p?(z); the di,$rrence p (z) - p?(z) is postfire o" zero. An analogous proposition applies to concave functions (z), which follows from the above, since - ( z) is convex. 6. Some suggested functions for /oiicfions. If, in the interval (p , p'), ÿ (z) is a connected function qni d o e s not decrease as z increases, and if ¡ r (z) is conveae in an interval, in theqnal takes place the inequality q K p (z) c ç ', /('r(z)) is also connected. In eflet, from it follows that / r("{') <r('*") t'°' ) <j(r(r(-)) + r(r(-))), which demonstrates the proposition. The following diagram 8 is similarly demonstrated: oonveze, growing eoncave, decroi8aante related related related ooncave conveze, decreasing conoave convex concave, growing conoave concave Snr lea functions connexes at les inégalités entra les valeurs moyennes. ' connexe, croissante° conca e, croissante connexe,décroissante concare, dëcroiasante ' . 191 connexe,décroiaaante concave, décroissante On a certain. convrxe function. We saw above that r, | z - z, | is convex in any interval that includes at least one points z, , z, , . . . .. This can be used to form a connected function whose right and left derivatives are di8erent at all points of a real, countable set, given in advance. Let e8and c , , c , ... an infinite sequence of positive numbers, z, , z, , ... an infinite sequence of real and bounded numbers, and suppose the series 2'o, convergent. It follows from the research into CANTOR's principle of condensation of singularities that L cook set out in Diff-LfiROTH (T/îeorie der Fu "lfioneii cirer verändertichen reellen Clrösse, § i o8*) that the function which is connected as we saw above, has a unique derivative everywhere, except at points z, , z, , z, , .... At these points, the function has one derivative function on the right and another on the left, and we have ÇJ (z,) -/ (z,) zc, . If, for example, for the set (z,), we choose the rational numbers in the interval (o , i) in a given order, /(z) will be connected in this interval, and will have a finite derivative at all irrational points, while at rational points the derivative functions ä right and ä left differ by a positive number. In closing, I can't resist adding a few remarks. It seems to me that the notion of "related function" is about as fundamental as these: positive function", increasing function". If I'm not mistaken in this, the notion should find its place in the elementary expositions of the theory of real functions. J. L. ÀI". À'. Jensen. As for the deÔnition o f a connected function rte }i1several variables the next is the most natural. The function rce1le p (X) alu roint illlil ÿtiqnc rcel1e N - (z , y, r, ...) is connected, if rlans a domain Simply connected and coiiveao we always have <-C -2 2" Il is + ' '. ' evirlent that such a tonc tion is always f'inction related to each of the xŒia1'1es. L'in i et se ri a pas lieu comme on le voit r-- un f(° ' '7J l *'* ' Additi0rt. After iivoir made the conference above ] I noticed that the fontlamental formula (3) was not entirely new as I thought. I've just found, 'fans une mémoire tle II. A . l'ltINGSI4EIM,' a quotation f r o m a note by II. 0. HOLIJ Eli ' rtans 1ar{ucl1e is de- shown the formula in question. A la vérite les li1'pot1ieses de Il. HöLDER are quite different from mine in that he assumes that 'r"(z) exists. Tva very important formula ( 5') is n o t ' mentioned more than the most of the -rl"'cations given above. En mcme temps je veux 'lemontrer une inégalité c1'un autre caractère que celles tlonnées plus haut. hans une addition1 a se mémoire précitée II. Then Gsiii iii astonishes an é1é¿-ante demonstration, which he attributes to M. LÜi:oa ir, 'le t h e inequality r'i the /' are positive and z > I. ' Zmr Theorie der petizeii Ii-anscendenten 4 "unktionen, S i t z u n gs b e r . d. m a t h . p li ya. G la s s e d. k. ba y er. A k a d. d . \V., t. 3.2, p. i 63 - 9* Nachtrag . . ibid. P 295 -5 3 ' Über einett. -Mittelicertsalz, fi a c h r. v. cl. k. G e se 11 s c h. d. V'. z u Göt t i ri g e n, On convex functions and inequalities between mean values. 193 This demonstration is easy to generate. L e t z be a positive variable and Ç(z) a positive function el increasing with z, and let o, , a, , . .. ô , , b , . . . of8 equally positive constants, we have 5(b,) < [ b) for r i , 2 , ... it, f' being the sum bl -[- b, -]- ... -]- ô,. By multiplying by a, the two members of t h e above inequality, and by Naming for r I , 2 , . . . n, we find on This is the inequality I had in mind. For o, b and ÿ (z) z '°', we find inequality (e). Using this inequality, we can generalize the conditions under which formula (8) is valid. Posing in e8and dana this n'q for a,q, we have which shows that formula (8) is still valid for only t h e equal signJ is to be discarded.