224-227
ISSN: 2222-2510
2011 WAP journal. www.waprogramming.com
Abstract: The average of ratios (r) is often approximated by the ratio of averages (q), but the reliability of
this depends on the distributions of numerator and denominator, neither of which is necessarily known.
Here bounds on their relative size (q/r) and difference (q r) are determined using Karamatas inequality.
Keywords: average inequality ratio
I.
INTRODUCTION
The statistical properties of two different averages of ratios formed from 0 < amin a1, a2, ..., an amax and 0
< bmin b1, b2, ..., bn bmax have been considered previously [1]. These variants are the ratio of averages
q=
i =1 i
n
b
i =1 i
a
b
(1)
The expected values (E(q) and E(r)), the variances (var(q) and var(r)) and the distributions of q and r are
known for uniformly distributed a and b [1]. From these it is clear that the ratio of the expected values is
E (q )
2(B 1)
=
<1,
E(r ) (B + 1)ln (B )
where B = bmax/bmin > 1, and the difference between the expected values is
(3)
B (B + 1)ln (B )
1
<0,
(4)
B + 1
2(B 1)
where A = amax/amin and L = amin/bmax [1]. While (3) and (4) may be helpful, they do not necessarily clarify the
situation because the distributions of a and b need to be known in order to understand the properties of q and r.
These distributions are not necessarily known and it is often difficult to determine them because the sample size
(n) is small. At least two questions are prompted by these observations: First, what are the limits of the relative
magnitudes of q and r? Second, what mathematical limits on the magnitudes of the difference between q and r
can be estimated?
E(q ) E(r ) = L( A + 1)
Here I use Karamatas inequality [3, 4] to determine bounds on q/r which are then used to determine bounds
on q r in terms of r or q.
II.
PRELIMINARY MATERIAL
I assume that 0 < amin a1, a2, ..., an amax and 0 < bmin b1, b2, ..., bn bmax and define L = amin/bmax, U =
amax/bmin, A = amax/amin and B = bmax/bmin. Karamatas inequality [3, 4] can be written
224
Simon Brown, World Applied Programming, Vol (2), No (4), April 2012.
K 2
i =1
i =1
where
K=
a b
n
i =1
ai
bi
B 1+ U L
B+ U L
K2,
) >1
(5)
(6)
In using (5) and (6), several other inequalities are also useful. First, as is well known (for example [5]),
(B + 1) ,
1
(7)
g (n )
b
4B
where Schweitzers inequality [6] gives the upper bound which can be improved slightly for odd n as shown by
Lupa [7]
2
1 b
g (n ) =
.
4B
8Bn 2
The lower bound can be obtained from the Cauchy-Schwarz inequality [8]. Second,
U
,
L
where the lower limit is in [3] and the upper limit is obtained from (6). Third
1< K2
L
4B
2
1,
U K (B + 1)2
where the upper limit is obvious since K2 > 1 [3] and B 1 and the lower limit is obtained from (6).
III.
(8)
(9)
(10)
It is obvious from (1) and (2) that 0 < L q U and 0 < L r U, so a trivial estimate of the bounds of q/r is
L
1
q
U
=
AB = .
(11)
U AB r
L
These bounds can be improved by substituting (2) into Karamatas inequality (5) and recalling that <a> = <b>q
(1). This yields
1 q
K2
b r
K 2 b
(12)
q
K2 ,
(13)
r
K (B + 1)
where I have neglected the second term of (8). It is easy to show that these bounds are at least as good as those in
(11) and that they are only equal (11) when B = 1 and A = amax/amin = 1 (Figure 1). Since lim B K 2 = A , the
bounds of (13) can be summarised as
2
L
L
4B
q
U
(14)
<4 2
K2 A = ,
U
U K (B + 1)2 r
L
as is illustrated in Figure 1. The upper limit of (13) follows directly from (9) and the lower limit follows directly
from (10). If B = 1 then K = 4B/(K2(B + 1)2) = q/r = 1.
225
Simon Brown, World Applied Programming, Vol (2), No (4), April 2012.
Figure 1. Comparisons of the bounds on q/r given by (11) ( ) and (13) ( ) with the ratio of expected values (3) () as a
function of B (A = 2). For clarity, the graph is truncated at the top and U/L = AB increases monotonically with B.
IV.
A trivial estimate of the bounds on the magnitude of q r can be obtained by analogy with (11) using the
bounds of q and r that follow from (1) and (2)
L U q r U L .
An initial estimate of improved bounds can be obtained from (13)
(15)
4B
2
2
(16)
K (B + 1)2 1r q r K 1 r
(Figure 2). From (10), the lower bound of (16) is better than that of (15), but the upper bound of (16) is better
than that of (15) only if r is small, so
4B
L U 2
1r q r f (r ) U L ,
2
K (B + 1)
(17)
Figure 2. Comparisons of the bounds on (q r)/L given by (15) ( ) and (16) ( ) with the difference of expected values (4) ()
as a function of B (A = 2). For clarity, the graph is truncated at the top and the bottom and |U L| increases monotonically with B.
226
Simon Brown, World Applied Programming, Vol (2), No (4), April 2012.
where
U L
2
K 1 L r K 2 1
f (r ) =
.
U L
U L
r> 2
K 1
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
S. Brown, "Averaging ratios: characteristics of the error of approximation," World Applied Programming, vol. 1, pp. 288-293, 2011.
T. M. Culley, L. E. Wallace, K. M. Gengler-Nowak, and D. J. Crawford, "A comparison of two methods of calculating GST, a genetic
measure of population differentiation," American Journal of Botany, vol. 89, pp. 460-465, 2002.
J. Karamata, "Ingalits relatives aux quotients et la diffrence de fg et f g," Publications de l'Institute Mathmatique (Beograd),
vol. 2, pp. 131-145, 1948.
A. Lupa, "On two inequalities of Karamata," Publikacije Elektrotehnickog Fakulteta Univerziteta u Beogradu - Matematika i Fizika,
vol. 617, pp. 119-123, 1978.
S. Simic, "On a converse of Jensen's discrete inequality," Journal of Inequalities and Applications, vol. 2009, pp. 153080, 2009.
P. Schweitzer, "Egy egyenltkensg az aritmetkai kzprtkrl," Mathematikai s Physikai Lapok, vol. 23, pp. 257-261, 1914.
A. Lupa, "A remark on the Schweitzer and Kantorovich inequalities," Publikacije Elektrotehnickog Fakulteta Univerziteta u Beogradu
- Matematika i Fizika, vol. 383, pp. 13-15, 1972.
D. S. Mitrinovi, J. E. Peari, and A. M. Fink, Classical and new inequalities in analysis, Dordrecht: Kluwer Academic Publishers,
1993.
227