On an inequality of N. Levinson

Abstract

Authors

T. Popoviciu
Institutul de Calcul

Keywords

Ky Fan inequality

Paper coordinates

T. Popoviciu, Sur une inégalité de N. Levinson, Mathematica (Cluj), 6(29) (1964), pp. 301-306 (in French)

PDF

About this paper

Journal

Mathematica Cluj

Publisher Name

Published by the Romanian Academy  Publishing House

DOI
Print ISSN

1222-9016

Online ISSN

2601-744X

google scholar link

??

Paper (preprint) in HTML form

Original text
Rate this translation
Your feedback will be used to help improve Google Translate

ON AN INEQUALITY BY N. LEVINSON

by
TIBERIU POPOVICIU
in Cluj

  1. 1.

    We will assume knowledge of the definition and main properties of divided differences. We will denote by [x1,x2,,xm+1;fx_{1},x_{2},\ldots,x_{m+1};fthe difference divided by ordermmof the functionffon the points, or the nodes,x1,x2,,xm+1x_{1},x_{2},\ldots,x_{m+1}The nodes may or may not be distinct. In cases where the nodes are not all distinct (or simple), the divided differences also involve successive derivatives of the function (over multiple nodes). In what follows, however, we will only consider divided differences over distinct nodes.

A function is said to be non-concave respectively(1)(\geqq-1)on the intervalII, if its difference divided by an angle of ordermmnon-negative respectively positive on all remainder distinct fromII.
em+2m+2points continue ifm1m\geq 1and has a rumblemmon an open interval is con-non-concave respectively conjunctionffis concave respectively conce d' is non-
2. The classical inequality of Jéproquement. (usual) vexes is as follows 2 with respect to conc- functions

f(α=1npαxαα=1npα)α=1npαf(xα)α=1npαf\left(\frac{\sum_{\alpha=1}^{n}p_{\alpha}x_{\alpha}}{\sum_{\alpha=1}^{n}p_{\alpha}}\right)\leqq\frac{\sum_{\alpha=1}^{n}p_{\alpha}f\left(x_{\alpha}\right)}{\sum_{\alpha=1}^{n}p_{\alpha}} (1)

and is verified regardless of the natural numbernn, the pointsxαIx_{\alpha}\in I,α=1.2,,n\alpha=1,2,\ldots,nand positive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,n, for any functionffnon-concave of order 1 on the intervalII.

Ifn>1n>1and if the functionffis convex of order 1 onIIThe equality in (1) holds if and only if the pointsxα,α=1.2,,nx_{\alpha},\alpha=1,2,\ldots,nare all lumped together.

Several proofs of these properties are known. Note that it suffices to prove inequality (1) in the case where the pointsxα,α==1.2,,nx_{\alpha},\alpha==1,2,\ldots,nare distinct, the functionffbeing non-concave of order 1 onIIand to show that ifn>1n>1Andffis convex of order 1 onII, then equality in (1) is not possible.

Eithern>1n>1and let's choose the notations so that we have
(2)

x1<x2<<xnx_{1}<x_{2}<\ldots<x_{n}

Ifpα>0,α=1.2,,np_{\alpha}>0,\alpha=1,2,\ldots,nand if we ask

ξα=p1x1+p2x2++pαxαp1+p2++pα,α=1.2,,n\xi_{\alpha}=\frac{p_{1}x_{1}+p_{2}x_{2}+\cdots+p_{\alpha}x_{\alpha}}{p_{1}+p_{2}+\cdots+p_{\alpha}},\quad\alpha=1,2,\ldots,n (3)

we then

ξ1=x1,ξn=α=1npαxαα=1npα And ξα<ξα+1<xα+1,α=1.2,,n1\xi_{1}=x_{1},\quad\xi_{n}=\frac{\sum_{\alpha=1}^{n}p_{\alpha}x_{\alpha}}{\sum_{\alpha=1}^{n}p_{\alpha}}\text{ and }\xi_{\alpha}<\xi_{\alpha+1}<x_{\alpha+1},\alpha=1,2,\ldots,n-1 (4)

Legality

α=1npαf(xα)f(ξn)α=1npα=\displaystyle\sum_{\alpha=1}^{n}p_{\alpha}f\left(x_{\alpha}\right)-f\left(\xi_{n}\right)\sum_{\alpha=1}^{n}p_{\alpha}= (5)
=α=1n1pα+1(p1+p2++pα)(ξαxα+1)2p1+p2++pα+1[ξα,ξα+1,xα+1;f]\displaystyle=\sum_{\alpha=1}^{n-1}\frac{p_{\alpha+1}\left(p_{1}+p_{2}+\ldots+p_{\alpha}\right)\left(\xi_{\alpha} -x_{\alpha+1}\right)^{2}}{p_{1}+p_{2}+\ldots+p_{\alpha+1}}\left[\xi_{\alpha},\xi_{\alpha+1},x_{\alpha+1};f\right]

then demonstrates Jensen's inequality (1).
We can state the properties expressed by this inequality also in the form of

THEOREM 1. For inequality (1) to be satisfied for all natural numbersnn, for any group ofnnpointsxαI,α=1.2,,nx_{\alpha}\in I,\alpha=1,2,\ldots,n, and for any group ofnnpositive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,nit is necessary and sufficient that the functionffeither non-concave of order 1 on the intervalII.

For the strict inequality (with the sign <) (1) to be satisfied for all natural numbersn>1n>1any group ofnnpointsxαIα=1.2,,nx_{\alpha}\in I\alpha=1,2,\ldots,n, not all of them together and any group ofnnpositive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,nit is necessary and sufficient that the functionffeither convex of order 1 on the intervalII.

The necessity of the conditions in the statement follows from formula (5), noting that ifx1<x2<x3x_{1}<x_{2}<x_{3}are any three points ofII(subject to the inequalities mentioned), it is always possible to find two positive numbersp,qp,qso thatx2=px1+qx3p+qx_{2}=\frac{px_{1}+qx_{3}}{p+q}(For example,p=x3x2,q==x2x1)p=x_{3}-x_{2},q=\left.=x_{2}-x_{1}\right)
3. Considernnpositive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,nand two groups, each ofnnpoints,xα,xα,α=1.2,,nx_{\alpha},x_{\alpha}^{\prime},\alpha=1,2,\ldots,nLet (3) and

ξα=p1x1+p2x2++pαxαp1+p2++pα,α=1.2,,n\xi_{\alpha}^{\prime}=\frac{p_{1}x_{1}^{\prime}+p_{2}x_{2}^{\prime}+\cdots+p_{\alpha}x_{\alpha}^{\prime}}{p_{1}+p_{2}+\cdots+p_{\alpha}},\alpha=1,2,\ldots,n (6)

Levinson's inequality is written

α=1npdf(xα)α=1npαf(α=1npαxαα=1npα)α=1npα(xα)α=1npαf(α=1npαxαα=1npα)\frac{\sum_{\alpha=1}^{n}p_{d}f\left(x_{\alpha}\right)}{\sum_{\alpha=1}^{n}p_{\alpha}}-f\left(\frac{\sum_{\alpha=1}^{n}p_{\alpha}x_{\alpha}}{\sum_{\alpha=1}^{n}p_{\alpha}}\right)\leqq\frac{\sum_{\alpha=1}^{n}p_{\alpha}\left(x_{\alpha}^{\prime}\right)}{\sum_{\alpha=1}^{n}p_{\alpha}}-f\left(\frac{\sum_{\alpha=1}^{n}p_{\alpha}x_{\alpha}^{\prime}}{\sum_{\alpha=1}^{n}p_{\alpha}}\right) (7)

where the pointsxα,xαx_{\alpha},x_{\alpha}^{\prime}and the functionffverify the following conditions. Suppose thatn>1n>1that we have (2) and that

x1+x1=x2+x2==xn+xn,xnxnx_{1}+x_{1}^{\prime}=x_{2}+x_{2}^{\prime}=\ldots=x_{n}+x_{n}^{\prime},x_{n}\leqq x_{n}^{\prime} (8)

We deduce from thisxn<xn1<<x1x_{n}^{\prime}<x_{n-1}^{\prime}<\ldots<x_{1}^{\prime}Andξα<ξα+1<xα+1\xi_{\alpha}<\xi_{\alpha+1}<x_{\alpha+1}\leqq
xα+1<ξα+1<ξα,α=1.2,,n1\leqq x_{\alpha+1}^{\prime}<\xi_{\alpha+1}^{\prime}<\xi_{\alpha}^{\prime},\quad\alpha=1,2,\ldots,n-1, and
(9)

ξαxα+1=xα+1ξα,α=1.2,,n1\xi_{\alpha}-x_{\alpha+1}=x_{\alpha+1}^{\prime}-\xi_{\alpha}^{\prime},\alpha=1,2,\ldots,n-1

The equalities (9) are equivalent ton1n-1equalities (8).
Note also that we have

[ξα,ξα+1,xα+1;f][ξα,ξα+1,xα+1;f]=\displaystyle{\left[\xi_{\alpha}^{\prime},\xi_{\alpha+1}^{\prime},x_{\alpha+1}^{\prime};f\right]-\left[\xi_{\alpha},\xi_{\alpha+1},x_{\alpha+1};f\right]=} (10)
=(ξαxα+1)[ξα,ξα+1,xα+1,ξα;f]+\displaystyle\quad=\left(\xi_{\alpha}^{\prime}-x_{\alpha+1}\right)\left[\xi_{\alpha},\xi_{\alpha+1},x_{\alpha+1},\xi_{\alpha}^{\prime};f\right]+
+(ξα+1ξα+1)[ξα,ξα+1,ξα,ξα+1;f]+\displaystyle\quad+\left(\xi_{\alpha+1}^{\prime}-\xi_{\alpha+1}\right)\left[\xi_{\alpha},\xi_{\alpha+1},\xi_{\alpha}^{\prime},\xi_{\alpha+1}^{\prime};f\right]+
+(xα+1ξα)[ξα,ξαξα+1,xα+1;f]\displaystyle\quad+\left(x_{\alpha+1}^{\prime}-\xi_{\alpha}\right)\left[\xi_{\alpha},\xi_{\alpha}^{\prime}\xi_{\alpha+1}^{\prime},x_{\alpha+1}^{\prime};f\right]

and it follows that, under the assumptionsn>1n>1, (2) and (8), we have[ξα,ξα+1,xα+1;f]\left[\xi_{\alpha}^{\prime},\xi_{\alpha+1}^{\prime},x_{\alpha+1}^{\prime};f\right]\geqqrespectively>[ξα,ξα+1,xα+1;f],α=1.2,,n1>\left[\xi_{\alpha},\xi_{\alpha+1},x_{\alpha+1};f\right],\alpha=1,2,\ldots,n-1depending on the functionffis non-concave respectively convex of order 2 on an interval containing the pointsxα,xα,α=1.2,,nx_{\alpha},x_{\alpha}^{\prime},\alpha=1,2,\ldots,n.

Taking into account (5) and (10), we have

[α=1npαf(xα)f(ξn)α=1npα][α=1npαf(xα)f(ξn)α=1npα]=\displaystyle{\left[\sum_{\alpha=1}^{n}p_{\alpha}f\left(x_{\alpha}^{\prime}\right)-f\left(\xi_{n}^{\prime}\right)\sum_{\alpha=1}^{n}p_{\alpha}\right]-\left[\sum_{\alpha=1}^{n}p_{\alpha}f\left(x_{\alpha}\right)-f\left(\xi_{n}\right)\sum_{\alpha=1}^{n}p_{\alpha}\right]=} (11)
=α=1npα+1(p1+p2++pα)(ξαxα+1)2p1+p2++pα+1\displaystyle=\sum_{\alpha=1}^{n}\frac{p_{\alpha+1}\left(p_{1}+p_{2}+\ldots+p_{\alpha}\right)\left(\xi_{\alpha}-x_{\alpha+1}\right)^{2}}{p_{1}+p_{2}+\ldots+p_{\alpha+1}}
{(ξαxα+1)[ξα,ξα+1,xα+1,ξα;f]++(ξα+1ξα+1)[ξα,ξα+1,ξα,ξα+1;f]++(xα+1ξα)[ξα,ξα,ξα+1,xα+1;f]}\displaystyle\cdot\left\{\begin{array}[]{l}\left(\xi_{\alpha}^{\prime}-x_{\alpha+1}\right)\left[\xi_{\alpha},\xi_{\alpha+1},x_{\alpha+1},\xi_{\alpha}^{\prime};f\right]+\\ +\left(\xi_{\alpha+1}^{\prime}-\xi_{\alpha+1}\right)\left[\xi_{\alpha},\xi_{\alpha+1},\xi_{\alpha}^{\prime},\xi_{\alpha+1}^{\prime};f\right]+\\ +\left(x_{\alpha+1}^{\prime}-\xi_{\alpha}\right)\left[\xi_{\alpha},\xi_{\alpha},\xi_{\alpha+1}^{\prime},x_{\alpha+1}^{\prime};f\right]\end{array}\right\}
  1. 4.

    From the preceding analysis, it follows that Levinson's inequality (7) is verified for all natural numbersnn, the pointsxα,xαIx_{\alpha},x_{\alpha}^{\prime}\in I,α=1.2,,n\alpha=1,2,\ldots,nsuch as one might have

{x1+x1=x2+x2==xn+xnmax(x1,x2,,xn)min(x1,x2,,xn)\left\{\begin{array}[]{l}x_{1}+x_{1}^{\prime}=x_{2}+x_{2}^{\prime}=\ldots=x_{n}+x_{n}^{\prime}\\ \max\left(x_{1},x_{2},\ldots,x_{n}\right)\leqq\min\left(x_{1}^{\prime},x_{2}^{\prime},\ldots,x_{n}^{\prime}\right)\end{array}\right.

and positive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,n, for any functionffnonconcave of order 2 on the intervalII.

Ifn>1n>1and if the functionffis convex of order 2 onIIThe equality in (7) holds if and only if the pointsxα,α=1.2,,nx_{\alpha},\alpha=1,2,\ldots,n(so also the pointsxα,α=1.2,,nx_{\alpha}^{\prime},\alpha=1,2,\ldots,n) are all confused.
N. LEVINSON obtained [1] the previous results, assuming that the functionffhas a third derivativef"f^{\prime\prime}We have therefore obtained a generalization of Levinson's inequality.

We can also state these properties in the form of
Theorem 2. For inequality (7) to be satisfied, for all natural numbersnn, for any group of2n2npointsxαxαIα=1.2,,nx_{\alpha}x_{\alpha}^{\prime}\in I\alpha=1,2,\ldots,n, such that we have (12) and for any group ofnnpositive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,nit is sufficient that the functionffand the function must continueffeither non-concave of order 2 on the intervalII.

For the strict inequality (with the sign <) (7) to be satisfied for all natural numbersn>1n>1, for any group of2n2npointsxα,xαIα=1.2,,nx_{\alpha},x_{\alpha}^{\prime}\in I\alpha=1,2,\ldots,n, such that one has (12), thexαx_{\alpha}not all lumped together and for every group ofnnpositive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,nit is sufficient that the functionffand the function must continueffeither convex of order 2 on the intervalII.

In the demonstration, the restriction that the pointsxα,α=1.2,,nx_{\alpha},\alpha=1,2,\ldots,nbeing distinct does not restrict generality since by virtue of (12), ifxα=xβx_{\alpha}=x_{\beta}We havexα=xβx_{\alpha}^{\prime}=x_{\beta}^{\prime}and vice versa.

The necessity of the conditions in the statement follows from formula (7). If we positn=2,x1=x,x2=x2=x+32h,x1=x+3h,p1=1,p2==2n=2,x_{1}=x,x_{2}=x_{2}^{\prime}=x+\frac{3}{2}h,x_{1}^{\prime}=x+3h,p_{1}=1,p_{2}==2, We haveξ2=x+h,ξ2=x+2h\xi^{2}=x+h,\xi_{2}^{\prime}=x+2h, the first member of formula (11) becomesΔh3f(x)=f(x+3h)3f(x+2h)+3f(x+h)f(x)==6h3[x,x+h,x+2h,x+3h;f]\Delta_{h}^{3}f(x)=f(x+3h)-3f(x+2h)+3f(x+h)-f(x)==6h^{3}[x,x+h,x+2h,x+3h;f]But, we know [2], that ifΔh3f(x)0\Delta_{h}^{3}f(x)\geqq 0respectively>0>0, for allx,x+3hI,h>0x,x+3h\in I,h>0, the functionffassumed to be continuous, is non-concave, respectively convex of order 2 onII.

Given the necessity of the theorem's condition, the hypothesis of the function's continuity can generally be replaced by a less restrictive hypothesis (measurability, being bounded onII, etc.).
5. The proof we gave above of Levinson's inequality is based on the fact that the successive means (3) and (6) of the pointsxαx_{\alpha}and pointsxαx_{\alpha}^{\prime}are formed with the same weights and
these points are linked by the equalities (9). In reality, to be able to use a formula such as (10), it is only sufficient that one has

(ξαxα+1)2=(xα+1ξα)2,α=1.2,,n1\left(\xi_{\alpha}-x_{\alpha+1}\right)^{2}=\left(x_{\alpha+1}^{\prime}-\xi_{\alpha}^{\prime}\right)^{2},\quad\alpha=1,2,\ldots,n-1 (13)

and not necessarily (9). By imposing on the pointsxα,xαx_{\alpha},x_{\alpha}^{\prime}The restriction (13), without all the equalities (9) being verified, allows us to obtain other theorems, analogous to Theorem 2.6.
As an example, suppose that we have

ξαxα+1=ξαxα+1,α=1.2,,n1\xi_{\alpha}-x_{\alpha+1}=\xi_{\alpha}^{\prime}-x_{\alpha+1}^{\prime},\quad\alpha=1,2,\ldots,n-1 (14)

assuming that the averagesξα,ξα\xi_{\alpha},\xi_{\alpha}^{\prime}are given by (3) and (6). We deduce that

x1x1=x2x2==xnxn=y.x_{1}^{\prime}-x_{1}=x_{2}^{\prime}-x_{2}=\ldots=x_{n}^{\prime}-x_{n}=y. (15)

If inequalities (2) are satisfied, we also havex1<x2<<xn,ξαξα=y,α=1.2,,nx_{1}^{\prime}<x_{2}^{\prime}<\ldots<x_{n}^{\prime},\xi_{\alpha}^{\prime}--\xi_{\alpha}=y,\alpha=1,2,\ldots,nand we alwaysξα<ξα+1<xα+1,α=1.2,\xi_{\alpha}<\xi_{\alpha+1}<x_{\alpha+1},\alpha=1,2,\ldots,,n1\ldots,n-1We have the formula.

[ξα,ξα+1,ξn+1;f][ξα,ξα+1,xα+1;f]=\displaystyle{\left[\xi_{\alpha}^{\prime},\xi_{\alpha+1}^{\prime},\xi_{n+1}^{\prime};f\right]-\left[\xi_{\alpha},\xi_{\alpha+1},x_{\alpha+1};f\right]=} (16)
=y{[ξα,ξα,ξα+1,xα+1;f]+[ξα,ξα+1,ξα+1,xα+1;f]+\displaystyle=y\left\{\left[\xi_{\alpha},\xi_{\alpha}^{\prime},\xi_{\alpha+1}^{\prime},x_{\alpha+1}^{\prime};f\right]+\left[\xi_{\alpha},\xi_{\alpha+1},\xi_{\alpha+1}^{\prime},x_{\alpha+1}^{\prime};f\right]+\right.
+[ξα,ξα+1,xα+1,xα+1;f]}\displaystyle\left.+\left[\xi_{\alpha},\xi_{\alpha+1},x_{\alpha+1},x_{\alpha+1}^{\prime};f\right]\right\}

and it follows that, under the assumptions (2), (15) (or (14)) andy>0y>0, 1a difference (10) (or (16)) still verifies the non-concavity property of order 2 mentioned above.

As above, we deduce
Theorem 3. For inequality (7) to be verified for all natural numbersnn, for any group of2n2npointsxα,xαI,α=1.2,,nx_{\alpha},x_{\alpha}^{\prime}\in I,\alpha=1,2,\ldots,n, such that one has (15) withy>0y>0and for any group ofnnpositive numberspα,α==1.2,,np_{\alpha},\alpha==1,2,\ldots,nit is sufficient that the functionffand it is necessary that the function be continuous (or measurable, or bounded etc.)ffeither non-concave of order 2 on the intervalII.

For the strict inequality (with the sign <) (7) to be satisfied for all natural numbersn>1n>1, for any group of2n2npointsxα,xαI,α=1.2x_{\alpha},x_{\alpha}^{\prime}\in I,\alpha=1,2,,n\ldots,n, THExαx_{\alpha}not all together, such as one has (15) withy>0y>0and for any group ofnnpositive numberspα,α=1.2,,np_{\alpha},\alpha=1,2,\ldots,nit is sufficient that the functionffand the continuous function (measurable, bounded, etc.) must be convex of order 2 on the intervalII.

To establish the necessity of the conditions in the statement, it suffices to taken=3n=3Andx1=x,x2=x1=x+h,x3=x2=x+2h,x3=x+3hx_{1}=x,x_{2}=x_{1}^{\prime}=x+h,x_{3}=x_{2}^{\prime}=x+2h,x_{3}^{\prime}=x+3hand inequality (7) becomes evenΔn3f(x)0\Delta_{n}^{3}f(x)\geqq 0withh>0h>07.
Inequality (7), under the hypotheses of Theorem 3 (II(being an open interval) can be demonstrated for a functionffnon-concave (respectively convex) of order 2, noting that this function has a
non-concave (respectively convex) derivative of order 1. By Jensen's inequality, the function ofyy,

α=1npαf(xα+y)f(ξn+y)α=1npα\sum_{\alpha=1}^{n}p_{\alpha}f\left(x_{\alpha}+y\right)-f\left(\xi_{n}+y\right)\sum_{\alpha=1}^{n}p_{\alpha}

is, non-decreasing (respectively increasing ifn>1n>1and thex1,x2,,xnx_{1},x_{2},\ldots,x_{n}do not all coincide), its derivative being non-negative (respectively positive).

BIBLIOGRAPHY
[1] Levinson N., Generalization of an Inequality of Ky Fan, Journal of Math. Analysis and Afplication, 8, 133, 134 (1964).
[2] Popoviciu T., On some properties of functions of one or two real variables, Mathematica, 8, 1-85 (1934).

Received on 12. III. 1963

1964

Related Posts