Notes on higher order convex functions (VII)

Abstract

Authors

Keywords

?

Paper coordinates

T. Popoviciu, Notes sur les fonctions convexes d’ordre supérieur (VII), Bull. de la Sect. Sci. de l’Acad. Roum., 22 (1939) no. 1, pp. 29-33 (in French).

PDF

About this paper

Journal

Bulletin de la Section Scientifique de l’Académie Roumaine

Publisher Name
DOI
Print ISSN
Online ISSN

[MR0001254, Zbl 0024.02201, JFM 65.0214.01]

google scholar link

??

Paper (preprint) in HTML form

BY

TIBERIU POPOVICIU

Note presented by Mr. S. Stoilov, MC. AR, in the session of July 14, 1939.

ON THE APPEARANCE OF ORDER FUNCTIONSnn

I. A functionf=f(x)f=f(x), defined and uniform on any linear set E is said to be convex, non-concave, polynomial, non-convex, or concave of ordernnfollowing that the inequality

[x1,x2,,xn+2;f]>,,=,Or<o\left[x_{1},x_{2},\ldots,x_{n+2};f\right]>,\geqq,=,\leq\mathrm{ou}<\mathrm{o}

is satisfied whateverx1,x2,,xn+2E1)\left.x_{1},x_{2},\ldots,x_{n+2}\in E^{1}\right).
All these functions are order functionsnn.
We will say that the set E is decomposed intommconsecutive subsets

E1,E2,,Em\mathrm{E}_{1},\mathrm{E}_{2},\ldots,\mathrm{E}_{m}

ifEiE,i=1.2,,m\mathrm{E}_{i}\subset\mathrm{E},i=1,2,\ldots,m, if every point of E belongs to an E and if every point ofEi\mathrm{E}_{i}is to the left of any point ofEi+1,i=1.2,,m1\mathrm{E}_{i+1},i=1,2,\ldots,m-1.

In our Thesis 2 ) we have demonstrated the following property:
Theorem I. Ifffis in ordernnon E, we can decompose this set into at mostk+1k+1consecutive subsets on each the function being of ordernknk.

The property is true fork=I,2,,n+Ik=I,2,\ldots,n+I.
This property follows from
Theorem II. Ifffis in ordernnand if
(I)

x1<x2<<xm,m>n+2x_{1}<x_{2}<\ldots<x_{m},m>n+2

is any tini subset of E , the sequence
(2)

Δnk+11,Δnk+12,,Δnk+1mn+1\Delta_{n-k+1}^{1},\Delta_{n-k+1}^{2},\ldots,\Delta_{n-k+1}^{m-n+1}
00footnotetext:1 ) For the notations and properties of order functionsnnsee our previous work.
2 ) Tiberiu Popoviciu, On some properties of functions of one or two real variables, Thesis, Paris 1933 or Mathematica 8, 1-85 (1934).

present at mostkkvariations of signs. Here we have posed
(3)Δki=[xi,xi+1,,xi+k;f],i=I,2,,mk,k=0,I,,mI\Delta_{k}^{i}=\left[x_{i},x_{i+1},\ldots,x_{i+k};f\right],i=\mathrm{I},2,\ldots,mk,k=0,\mathrm{I},\ldots,m-\mathrm{I}

In this note we propose to give a converse of Theorem II.
2. Before going further let us make some remarks on the order functionsnn. Using notation (3), any divided difference of ordern+In+Iofffonn+2n+2points, chosen from points (I), is an arithmetic mean of the divided differencesΔn+11,Δn+12,,Δn+1mn1\Delta_{n+1}^{1},\Delta_{n+1}^{2},\ldots,\Delta_{n+1}^{mn-1}, SO

[xi1,xi2,,xin+2;f]=i=1mn1HASiΔn+1iHASi0,i=1.2,,mn1,i=1mn1HASi=1\begin{gathered}{\left[x_{i_{1}},x_{i_{2}},\ldots,x_{i_{n+2}};f\right]=\sum_{i=1}^{mn-1}\mathrm{\penalty 10000\ A}_{i}\Delta_{n+1}^{i}}\\ \mathrm{\penalty 10000\ A}_{i}\geq 0,i=1,2,\ldots,mn-1,\sum_{i=1}^{mn-1}\mathrm{\penalty 10000\ A}_{i}=1\end{gathered}

THEHASi\mathrm{A}_{i}being independent of the functionff. We have surely
HASi>0,HASin+2nI>0\mathrm{A}_{i}>0,\mathrm{\penalty 10000\ A}_{i_{n+2}-n-\mathrm{I}}>0ifi1<i2<<in+2i_{1}<i_{2}<\ldots<i_{n+2}.
We immediately deduce
Lemma I. The necessary and sufficient condition for the functionff, defined on the finite set (1), either convex, non-concave, polynomial, nonconvex or concave of ordernnis that we have

Δn+1i>,,=,Or<o,i=I,2,,mnI\Delta_{n+1}^{i}>,\geq,=,\leqq\mathrm{ou}<\mathrm{o},i=\mathrm{I},2,\ldots,mn-\mathrm{I}

From this lemma immediately follows Theorem II.
Let us prove Lemma II again
. For the functionffeither of ordernnon E having at leastn+3n+3points it is necessary and sufficient that one has

[x1,x2,,xn+2;f][x2,x3,,xn+3;f]0\left[x_{1},x_{2},\ldots,x_{n+2};f\right]\left[x_{2},x_{3},\ldots,x_{n+3};f\right]\geq 0

whatever the pointsx1<x2<<xn+3x_{1}<x_{2}<\ldots<x_{n+3}of E.
The condition is obviously necessary. Let us show that it is also sufficient. It is sufficient to show that the property is not true for a function which is not of ordernn, it is therefore sufficient to demonstrate the

Lemma III. If the function f is not of ordernnon E we can findn+3n+3pointsx1<x2<<xn+3x_{1}<x_{2}<\ldots<x_{n+3}of E such that we have

[x1x2,,xn+2;L][x2,x3,,xn+3;f]<o\left[\begin{array}[]{llll}x_{1}&x_{2}&,\ldots,&x_{n+2}\end{array};l\right]\left[x_{2},x_{3},\ldots,x_{n+3};f\right]<o

The fact thatffis not in ordernnmeans that we can findn+2n+2pointsα1,α2,,σn+2\alpha_{1},\alpha_{2},\ldots,\sigma_{n+2}of E such that[α1,α2,,αn+2;f]>0\left[\alpha_{1},\alpha_{2},\ldots,\alpha_{n+2};f\right]>0Andn+2n+2pointsβ1β2,,βn+2\beta_{1}\beta_{2},\ldots,\beta_{n+2}of E such that[β1,β2,,βn+2;f]<0\left[\beta_{1},\beta_{2},\ldots,\beta_{n+2};f\right]<0Let's put all the separate points in order.oi,βio_{i},\beta_{i}in an increasing sequence (I). We then haven+3m2n+2n+3\leq m\leq 2n+2. The continuation of the divided differences

Δn+11,Δn+12,,Δn+1mn1\Delta_{n+1}^{1},\Delta_{n+1}^{2},\ldots,\Delta_{n+1}^{mn-1} (4)

has at least one sign variation, therefore contains at least two non-zero terms with opposite signs. LetΔn+1p\Delta_{n+1}^{p}the first non-zero term in (4) andΔn+1s\Delta_{n+1}^{s}the first non-zero term and of opposite sign withΔn+1p\Delta_{n+1}^{p}. Finally, be itΔn+1r\Delta_{n+1}^{r}the last non-zero term in the sequenceΔn+11,Δn+12\Delta_{n+1}^{1},\Delta_{n+1}^{2}, ,Δn+1s1\Delta_{n+1}^{s-1}. We then haveΔn+1r.Δn+1s<0\Delta_{n+1}^{r}.\Delta_{n+1}^{s}<0Andr<sr<s. Ifs=r+1s=r+1inequalityΔn+1rΔn+1r+1<o\Delta_{n+1}^{r}\Delta_{n+1}^{r+1}<\mathrm{o}proves Lemma III. Ifs>r+Is>r+\mathrm{I}we haveΔn+1r+1=Δn+1r+2==Δn+1s1=0\Delta_{n+1}^{r+1}=\Delta_{n+1}^{r+2}=\ldots=\Delta_{n+1}^{s-1}=0and inequality
[xr,xr+1,,xr+n,xs+n;f][xr+1,xr+2,,xr+n,xs+n,xs+n+1;f]<0\left[x_{r},x_{r+1},\ldots,x_{r+n},x_{s+n};f\right]\left[x_{r+1},x_{r+2},\ldots,x_{r+n},x_{s+n},x_{s+n+1};f\right]<0proves Lemma III.
3. Let us now establish the converse of Theorem II. To do this, let us note that[x1,x2,,xn+2;P]=o\left[x_{1},x_{2},\ldots,x_{n+2};\mathrm{P}\right]=\mathrm{o}identically if P is a polynomial of degreen1n^{1}). From this results this important property that any functionfPf-\mathrm{P}, where P is a polynomial of degreenn, enjoys the same property asffwith respect to any convexity character of ordern\geq n. We then have the

Theorem III. If, whatever the polynomial P of degreennand the finite subset ( I ) of E , the sequence (2) corresponding to ( I ) and to the functionfPf-\mathrm{P}present at mostkkvariations of signs, the functionffis in ordernnon E.

It is enough to demonstrate that if the function is not of ordernn, we can find a sequence (I) of E and a polynomial P such that the sequence (2) has more thankkvariations. Let us take for this, as a continuation (I),n+3n+3pointsx1<x2<<xn+3x_{1}<x_{2}<\ldots<x_{n+3}of E such thatΔn+11.Δn+12<0\Delta_{n+1}^{1}.\Delta_{n+1}^{2}<0, which is possible according to Lemma III. Let us first determine the polynomialP=has0xn+has1xn1++hask1λnk+1\mathrm{P}=a_{0}x^{n}+a_{1}x^{n-1}+\ldots+a_{k-1}\lambda^{n-k+1}so that for the functionf1=fPf_{1}=f-\mathrm{P}we have

Δnk+1i=0,i=2.3,,k+1\Delta_{n-k+1}^{i}=0,i=2,3,\ldots,k+1 (5)

The corresponding sequence (2) becomes, up to positive factors,

(I)kΔn+I1,0.0,,0k,Δn+I2.(-\mathrm{I})^{k}\Delta_{n+\mathrm{I}}^{1},\overbrace{0,0,\ldots,0}^{k},\Delta_{n+\mathrm{I}}^{2}. (6)

But (5), regarded as a system ofkklinear equations in thekkunknownshas0,has1,,hask1a_{0},a_{1},\ldots,a_{k-1}, has its determinant different from zero 2 ). We can therefore modify these coefficients so that the sequence (6) has all its terms non-zero and with alternating signs, therefore it presentsk+Ik+Ivariations of signs. Theorem III is therefore demonstrated. We see that in the statement we can only consider the finite subsets of E havingn+3n+3points.

00footnotetext: 1 ) of effective degreen\leq n.
)2\left.{}^{2}\right)This determinant is in fact equal, up to the sign, toi=2k[I=0ki(xn+2Ixi)](=1\prod_{i=2}^{k}\left[\prod_{j=0}^{k-i}\left(x_{n+2}-j-x_{i}\right)\right](=1Fork=1)k=1).

4. We have said that Theorem I follows from Theorem II. Theorem I also has a converse which we will study in another work. Let us make here only a remark on the casen=In=I. We demonstrated in a previous note 1 ) that inequality
(7)f(x2)max[f(x1),f(x3)],x1<x2<x3,x1,x2,x3E\quad f\left(x_{2}\right)\leq\max\left[f\left(x_{1}\right),f\left(x_{3}\right)\right],x_{1}<x_{2}<x_{3},x_{1},x_{2},x_{3}\in\mathrm{E}
is necessary and sufficient so that we can decompose the set E into at most two consecutive subsets such that on each the functionffOr -ffis monotonic, the monotonicity being in opposite directions on the two subsets. In particular, the first-order functions verify this property and are therefore such thatffOr -ffverifies inequality (7). Here we have again a reciprocal and thus we can state the

Theorem IV. For the functionffis of order I on E it is necessary and sufficient that, whatever the number a, we can decompose the set E into at most two consecutive subsets on each the functionfαxf-\alpha xbeing monotonic, the monotony being in opposite directions on the two subsets.

The condition is necessary sincefαxf-\alpha xis of order i ifffis of order I. Let us show that it is also sufficient. It suffices to show that ifffis not in orderIIwe can find a numberα\alphaso thatf1f_{1}Andf1-f_{1}, withf1=fαxf_{1}=f-\alpha x, do not verify the property expressed by inequality (7). Ifffis not of order I we can find 4 pointsx1x_{1}of<x3<x4<x_{3}<x_{4}of E so that

[x1,x2,x3;f][x2,x3,x4;f]<0\left[x_{1},x_{2},x_{3};f\right]\left[x_{2},x_{3},x_{4};f\right]<0

Let's determine the numberα\alphaso that.
[x2,x3;f]>has>max([x1,x2;f],[x3,x4;f])\left[x_{2},x_{3};f\right]>a>\max\left(\left[x_{1},x_{2};f\right],\left[x_{3},x_{4};f\right]\right)if[x1,x2,x3;f]>0[x2,x3;f]<α<min([x1,x2;f],[x3,x4;f])if[x1,x2,x3;f]<0\left[x_{1},x_{2},x_{3};f\right]>0\left[x_{2},x_{3};f\right]<\alpha<\min\left(\left[x_{1},x_{2};f\right],\left[x_{3},x_{4};f\right]\right)\operatorname{si}\left[x_{1},x_{2},x_{3};f\right]<0.

By posing thenf1=fαxf_{1}=f-\alpha x, we have

f1(x2)>f1(x1),f1(x3)<f1(x2),f1(x4)>f1(x3)f_{1}\left(x_{2}\right)>f_{1}\left(x_{1}\right),f_{1}\left(x_{3}\right)<f_{1}\left(x_{2}\right),f_{1}\left(x_{4}\right)>f_{1}\left(x_{3}\right)

in the first case and

f1(x2)<f1(x1),f1(x3)>f1(x2),f1(x4)<f1(x3)f_{1}\left(x_{2}\right)<f_{1}\left(x_{1}\right),f_{1}\left(x_{3}\right)>f_{1}\left(x_{2}\right),f_{1}\left(x_{4}\right)<f_{1}\left(x_{3}\right)

in the second case. We immediately check thatf1f_{1}Andf1-f_{1}do not verify inequality (7).

We can choose the numbersα,β\alpha,\betaso that iff1=f+αx+βf_{1}=f+\alpha x+\betawe havef1(x2)=HAS,f1(x3)=B,HAS,Bf_{1}\left(x_{2}\right)=\mathrm{A},f_{1}\left(x_{3}\right)=\mathrm{B},\mathrm{A},\mathrm{B}being any two numbers. Taking A negative, B positive sufficiently small if[x1,x2,x3;f]>o\left[x_{1},x_{2},x_{3};f\right]>\mathrm{o}and A positive, B negative sufficiently small if[x1,x2,x3;f]<0\left[x_{1},x_{2},x_{3};f\right]<0, we have

f1(x1)>0,f1(x2)<0,f1(x3)>0,f1(x4)<0f_{1}\left(x_{1}\right)>0,f_{1}\left(x_{2}\right)<0,f_{1}\left(x_{3}\right)>0,f_{1}\left(x_{4}\right)<0
00footnotetext: 1 ) Tiberiu. Popoviciu, Two remarks on convex functions, Bulletin Scient. Acad. Roumaine, 20, 45-49 (1938).

in the first case and

f1(x1)<0,f1(x2)>0,f1(x3)<0,f1(x4)>0f_{1}\left(x_{1}\right)<0,f_{1}\left(x_{2}\right)>0,f_{1}\left(x_{3}\right)<0,f_{1}\left(x_{4}\right)>0

in the second case. We deduce
Theorem V. For the function j to be of order I on E it is necessary and sufficient that, whatever the numbersα,β\alpha,\beta, we can decompose the set E into at most three consecutive subsets on each the functionf+αx+βf+\alpha x+\betabeing of invariable sign.

We can now glimpse the converse of Theorem I in the general case but, as we have said, we will return to this question in another paper.

Cernăuți, July 5, 1939.

1939

Related Posts