Probl. Anal. Issues Anal. Vol. 6(24), No. 1, 2017, pp. 3-10
DOI: 10.15393/j3.art.2017.3750
3
UDC 517.52
VANDANJAV ADIYASUREN, TSERENDORJ BATBOLD
EXTENSION OF THE REFINED GIBBS' INEQUALITY
Abstract. In this note, we give an extension of the refined Gibbs' inequality containing arithmetic and geometric means. As an application, we obtain converse and refinement of the arithmetic-geometric mean inequality.
Key words: arithmetic-geometric mean inequality, Jensen's inequality, log-function, Gibbs' inequality
2010 Mathematical Subject Classification: 26D15, 94A15
1. Introduction. Let n — 2 and wi,...,wn be non-negative real numbers such that Y^n=1 Wj = 1. Let An and Gn denote the weighted arithmetic and geometric means of the positive real numbers xi,... ,xn, that is,
An = ^^ Wj Xj and Gn = JJ x
wj j
j=i j=i
The arithmetic-geometric mean inequality asserts that
A > G
■^n — ^ n*
For more details about the arithmetic-geometric mean inequality the reader is referred to [3]—[8].
It is interesting that some classical inequalities such as arithmetic-geometric mean inequality (see [9]), the Jensen inequality (see [10]), the Holder inequality (see [11]) play an important role in information sciences.
Let Pj ,Qj > 0 (j = 1..., n) and ¿n=i pj = j j. Then
n
0 < £ Pj log pj j=i Qj
(Petrozavodsk State University, 2017
with equality if and only if pj = qj (j = 1,...,n). This inequality is known in literature as the Gibbs' inequality (see [8, p. 382]). The Gibbs' inequality has many applications in information theory and also in mathematical statistics.
In 2004 Halliwell and Mercer [6] presented the following refinement of the Gibbs inequality:
Theorem 1. Let pj ,qj (j = 1,... ,n) be positive real numbers satisfying
En v yn m
j=i Pj = £ J=i qj ■ Then
y qj - pj)2 < ypj log Pj < yqj (? - Pj)2 (1) j=1 qj + Mj < j=Pj g qj < j= q? + m ()
where mj = min(pj, qj), Mj = max(p?, q?) (j = 1,..., n).
In 2014 H. Alzer [2] proved the following refinement of (1): Theorem 2. Let a, 3 G R. Then, inequalities
n
qj (qj - pj) „ ^ p log qj (qj- pj) (2) £ j+mj < h Pj log qj < h q?+m j> (2)
hold for positive real numbers pj, qj (j = 1,... ,n) withJ2 jj=]_ Pj = £ jj=i. qj if and only if a < 1/3 and 3 > 2/3.
In this note we give an extension of (2) containing arithmetic and geometric means. As an application, we obtain refinement of the left-hand inequality in (1) and we also give a converse of the arithmetic-geometric mean inequality.
2. Main results. In order to prove our main results, we need the following lemmas.
Lemma 1. (see [2]) (i) If 0 < x < 1, then
(x - 1)2 (x - 1)2
x - 1 ---^ < log x < x - 1 ---f- (3)
x + x1/3 " " x + 1 v 7
with equality if and only if x = 1. (ii) If x > 1, then
1 (x - 1)2 ! 1 (x - 1)2 m
x - 1 - ---- < log x < x - 1 - ---4rr. (4)
x + 1 x + xi/3
-r f / \ (^P aa) _ riii с
Lemma 2. Let fa(x) := -¡-^—+ logx, a > 0. Then fa
a(x2 + max{x2, а2 })
is a concave function on (0,
Proof. If x > а, then fa (x) = ^X-OX—+ log x, and consequently,
fa (x) = ^ < 0.
x(x — a)2
a(x2 + a2) yields
On the other hand, if 0 < x < а, then fa (x) = 9-+ log x, which
22
„,,, N a6 + 7a4x2 — 9a2x4 + x6 (a2 — x2)(a4 +8a2x2 — x4)
J (x) =--= — --—-- < 0
Ja (X) x2(x2 + a2)3 x2(x2 + a2)3 <
Therefore, the function fa (x) is concave for x > 0. □
Lemma 3. (see [1]) Let fa be as defined in Lemma 2, k E {2,..., n — 1}, and
k
k \ /J2 w,j x,j \ k
wj fA J j=1k- — ^ WW fAn (x,j )
'j=1 J v Ew,^ j=1
j=1
Sk := max
Then,
0 < S2 < S3 < ■ ■ ■ < Sn-1. First, we give an extension of (2) based on the corresponding result in
[2].
Theorem 3. Let a, ^ E R and mj = min(x2, A^, Mj = max(x2, An) (j = 1,..., n). Then inequalities
1 ^ wj xj (xj — An)2 ^ , , i n ^ 1 ^ wj xj (xj — An)2
An x2 + mj <log An — log Gn < An 2=i j^fMp"
(5)
hold if and only if a < 1/3 and ^ — 2/3. Proof. We follow the method of proof given in [2].
(Necessity) Since the sums on the left-hand side and on the right-hand side of (5) are increasing with respect to a and respectively, it suffices
to prove (5) for a = 1/3 and 3 = 2/3. Therefore, substituting xj/An instead of x in (3) and (4), then multiplying by Wj, and summing, we obtain
^L Y^ I Wjxj (xj - An)2
A I W" xj W"A" 2 . 2/3, ,1/3
1 v^ f zi Wj (xj - An)2
= A" ^ ^j xj - ^j--^ 1/3,2/3 1 <
" x3- <An \ xj + xj A" ,
t x, 1 / . Wj(x?- - A")2
< Z^ W" log 4" < T 2s W"xj - W"A"--jA-
A" A" xj + A"
1 V^ /„.. ™ ... 4 Wjxj (xj A")2
1 \ ^ I A Wj x, (x,
= ^ z^ I x" - a" "j j j
A" ^ \ " " " " x2 + 1/2 Mi/2
" xj < A„ \ xj + mj Mj
and
^L Y^ I A Wjxj (xj - A")2
4 Wj xj Wj A" 2 , i/2,,i/2 A"xj >A„ V x2 + mj Mj'
1 \—^ f a Wj (x, — An)2\ x—^ x,
= ( Wj xj - Wj A"-------- I < Wj log A- <
A" x. >A V x j +A" / x. >A A"
1 I A Wj (xj - A")2
< — I Wjxj - WjA" 1/^2/3
" x3- >An \ xj + xj A"
= ^L Y^ I - A Wjxj (xj - A")2
= A Wj xj Wj A" 2 , 1/3,,2/3
A"xj >A„ V xj + .
Further, utilizing inequalities m2/3MjL/3 < mi/2MjL/2 < mi/3Mj2/3, we get
1 \ ^ / w j x j (x j A") \ ^ x j A w"x" - w"a" - 2 i 2/^,1/J < Z^ w" log a" <
A"xj <aA x2 + Mj- / Xj < A A"
1 / _ A _ Wjxj (xj - A")2
< A Z^ Wj x" Wj A" 2 , 1/3,,2/3
A"xj <A„ V x2 + m/ M/
and
1 \ ^ ( A —jXj (xj An)2 \ ^ "v Xj
^ £ -X - - An - 2j 2/3 „1/3 < £ - lQg ~T <
1 I , wj xj (xj — An )2
<A- £ (- Xj - Wj An - -jj n)
An x>A \ j X2 + m1/3M2/3
x— >A„ \ Xj + mj Mj
Combining this together, we obtain
1 ^^ Wj Xj (Xj - An)2 ________1 ^^ Wj Xj (Xj — An)2
-2X+ j An;3 < log An - !og Gn < f£ 2 + 1/3 M2/3.
An j=1 X2 + m/ M/ An j=1 x2 + m/ Mj'
j j j=i j j
as desired.
(Sufficiency) Let s, t E R with 1 < t < s + 1. Set s +1 - t
xi =-, x2 = t, Xj = 1 (j = 3,..., n);
s1
Wi = ---7--7-, W2 = -----—, Wj =
s + 1 + (n - 2)t' s + 1 + (n - 2)t' j s + 1 + (n - 2)t
(j = 3,..., n). Now, the same computation as in the proof of Theorem 2 (see [2]), provides that a - 1/3 and p > 2/3. □
Remark. Putting Xj = —, Wj = pj /^ "=i Qj (j = 1,...,n), where
j
E"=i Pj = sn=i Qj; in (5), we get (2).
By using Theorem 3, we obtain the following consequence. Theorem 4. The inequality
An - Gn <V j (j (6)
X2 + mßj
holds for p > 2/3.
Proof. By the mean value theorem and Theorem 3, we have
An - Gn = exp(log An) - exp(log Gn) < An (log An - log Gn) <
< ^ Wj Xj (Xj - An)2
- j=i X2 + mf Mj^ ,
t
which completes the proof. □
Our next result refines sign of the first inequality in (5) for a = 0.
Theorem 5. If C = V Wj- A")2 , then
A" j= x2 + Mj ,
C < C + S2 < C + S3 <•••< C + S"_i < log(A") - log(G"). (7)
Equality occurs if and only if all x, are equal. Proof. By Lemma 3 we have
C < C + S2 < C + S3 <•••< C + S"_i.
Now, we have to prove the last inequality in (7). Let's choose an arbitrary xw G {ii,...,x„}, 1 < < < ••• < < n, with the corresponding weights Wj G {wl, ..., W"}, and let xin = {xL,..., x„} \ {x^,... }. Now, utilizing the first inequality in (5) with a = 0, we obtain
log(A") = log
' /y Wj) "j I >
ii_x ii_ I
jn in
E"_1
v - / j=1 Wlj
^ 1 Wjn xjn (xjn A") + A" xjn + max(xjn, A")
— 1 / — 1
EI— w,j I En=—i1 w,j A"
A" n—1 w x 2 n—1 w x 2
"jwrj + max I,A"
/ n — 1
f fr"-1 W x ^ j ^ -log xwr x|j
Ere_i
x . j = 1 Wlj
1 ^ Wj xj (xj - A" )2 1 "_1 Wjj x^j (x^j - A" )2 + A" j=1 x2 + max(x2 A") A" j=1 x|j + max(x|j , A")
2
n— 1
+ log(Gn) - £ log
XUj +
j=1
= log(Gn) + C +
n-1
£ wU; I \j=1
n-1
£ wUj \j=1
j=1 wUj XUj
wU; 1 fAn
En— 1 j = 1 xuj
En— 1
j = 1 wuj xuj
En— 1
j — 1 wUj n1
En — 1
.7 = 1 w
(xUj ) •
j = 1
Since , i = {1,...,k} are arbitrary, the last inequality in (7) holds. The theorem is proved. □
Putting xj = pj,wj = pj/En=1 qj (j = where ^n=1 p, =
n j=
= n=1 qj, in Theorem 5, we obtain the following refinement of the first inequality in (1).
Corollary 1. Let pj, qj (j = 1,..., n) be positive real numbers satisfying
En v yn m
,-=i Pj = E 7=i qj • Then
n f \2 n
£ qj (qj - Pj) = C < (7 + S2 < (7 + S3 <■■■< C + Sn—1 < V Pj log j=1 q2 + M j=1
Pj qj
(8)
where
Sk = max
1<Ui<U2<^"<Ufc <n
^ ^ ^E7 = 1 qUj
£pw I f1 1 k
_ \j = 1 / \^j = 1 qU
£ Pu, /1 (^ j VPuj
and Mj = max(p2, qj2) (j = 1,..., n).
j
Acknowledgment. The authors wish to express their thanks to Professor Mario KrniC for his valuable suggestions. This work was supported by the Asia Research Center at the National University of Mongolia and the Korea Foundation of Advanced Studies (Project No. 18, 2016-2017).
References
[1] Adiyasuren V., Batbold Ts., Adil Khan M. Refined arithmetic-geometric mean inequality and new entropy upper bound. Commun. Korean Math. Soc., 2016, vol. 31, no. 1, pp. 95-100. DOI: 10.4134/CKMS.2016.31.1.095
[2] Alzer H. On an inequality from information theory. Rend. Istit. Mat. Univ. Trieste, 2012, vol. 46, pp. 231-235.
[3] Beckenbach E. F., Bellman R. Inequalities. Springer Verlag, Berlin, 1983.
[4] Bullen P. S., MitrinoviC D. S., VasiC P. M. Means and their inequalities. Reidel, Dordrecht, 1988.
[5] Gao P. A new approach to Ky Fan-type inequalities. Int. J. Math. Math. Sci., 2005, no. 22, pp. 3551-3574. DOI: 10.1155/IJMMS.2005.3551
[6] Halliwell G. T., Mercer P. R. A refinement of an inequality from information theory. J. Inequal. Pure Appl. Math., 2004, vol. 5, no. 1, pp. 1-3.
[7] Hardy G. H., Littlewood J. E., Polya G. Inequalities. Cambridge University Press, Cambridge, 1954.
[8] Mitrinovic D. S. Analytic inequalities. Springer Verlag, New York, 1970.
[9] Parkash O., Kakkar P. Entropy bounds using arithmetic-geometric-harmonic mean inequality. Int. J. Pure Appl. Math., 2013, vol. 89, no. 5, pp. 719-730. DOI: 10.12732/ijpam.v89i5.8
[10] Simic S. Jensen's inequality and new entropy bounds. Appl. Math. Lett., 2009, vol. 22, pp. 1262-1265. DOI: 10.1016/j.aml.2009.01.040
[11] Tian J. New property of a generalized Holder's inequality and its applications. Information Sciences, 2014, vol. 288, pp. 45-54. DOI: 10.1016/j.ins.2014.07.053
Received October 16, 2016. In revised form, January 30, 2017. Accepted January 30, 2017. Published online April 5, 2017.
National University of Mongolia
P.O. Box 46A/104, Ulaanbaatar 14201, Mongolia
E-mail: VAdiyasuren@yahoo.com, tsbatbold@hotmail.com