You are on page 1of 3

ANOTHER DISTANCE FUNCTION

AN HOA VU

Let f (t) = t log t. For a, b > 0, define


s  
f (a) + f (b) a+b
d(a, b) = −f
2 2
then d(a, b) is a metric on R+ .

Proof. First, d is well-defined because f (t) is a convex function i.e. f 00 (t) = 1


t >0
for t > 0.
The main thing to prove is the triangle inequality. Observe that
s  
a log a + b log b a+b a+b
d(a, b) = − log
2 2 2
s  
1 a+b
= a log a + b log b − (a + b) log
2 2
s     
1 a+b a+b
= a log a − log + b log b − log
2 2 2
s  
1 2a 2b
= a log + b log
2 a+b a+b
s  
a+b a 2a b 2b
= log + log
2 a+b a+b a+b a+b
r  
a+b a
= K
2 a+b
where the function (0, 1) → R≥0 is given by
r
p f (2t) + f (2 − 2t)
K(t) := t log(2t) + (1 − t) log(2(1 − t)) = .
2
(Check that K(t) is well-defined and K(t) is decreasing on [0, 1/2] and increasing
on [1/2, 1].)
Now the triangle inequality
d(a, b) ≤ d(a, c) + d(b, c)
r   r   r  
a+b a a+c c c+b c
⇐⇒ K ≤ K + K
2 a+b 2 a+c 2 c+b
  r   r  
a a+c c c+b c
⇐⇒ K ≤ K + K
a+b a+b a+c a+b c+b
1
2 AN HOA VU

a
Notice that the rational functions a+b , a+c
a+b , ... appearing on the last inequality
are all homogeneous of degree 0. So the triangle inequality is equivalent to
  r   r  
a a+1 1 1+b 1
(1) K ≤ K + K
a+b a+b a+1 a+b 1+b
for all a, b > 0 obtained by normalizing c = 1. (If (1) is true, applying it with a/c
and b/c in place of a and b yields the original inequality.)
We do one more normalization:
√ √ √
     
a 1 1
(1) ⇐⇒ a + b K ≤ 1+aK + 1+bK
a+b a+1 1+b
s  !
√ √
    
b 1 1 1
⇐⇒ a 1+ K ≤ 1+aK + 1+bK
a 1 + ab a+1 1+b

⇐⇒ a k(b/a) ≤ k(a) + k(b)
where the function k : R≥0 → R≥0 is given by

 
1
k(s) := 1 + s K
1+s

r
1 2 s 2s
= 1+s log + log
1+s 1+s 1+s 1+s
r
2 2s
= log + s log .
1+s 1+s
As
2s
log 1+s
k 0 (s) =
2k(s)
0 2s
the sign of k (s) is the same as the sign of log 1+s and so the function k(s) is
increasing on [1, +∞) and decreasing
√ on (0, 1).

Note that a k(b/a) = b k(a/b) so without loss of generality, we can assume
a < b. Changing variable x = b/a > 1 and so b = ax, one further has

(1) ⇐⇒ a k(x) ≤ k(a) + k(ax)
k(x) k(a) k(ax)
⇐⇒ √ ≤ √ + √
x ax ax
for all a > 0 and x > 1. From the second form, the inequality is true if a ≥ 1
because then x < ax and similar to the function k(s), the function s 7→ k(s)

s
is
k(x) k(ax)
increasing on [1, +∞) and decreasing on (0, 1] so we even have √
x
≤ √
ax
. Note
k(s) k(s)2
that we shouldn’t check √
s
is increasing: check instead.
s
Now let a ∈ (0, 1) be arbitrary. Observe that:
s
√ √
 
1 2 1 2/s
sk = s log + log
s 1 + 1/s s 1 + 1/s
r
2s 2
= s log + log
s+1 s+1
= k(s)
ANOTHER DISTANCE FUNCTION 3

so if 1 < x ≤ a1 , we also have a stronger inequality


√ √
 
1
a k(x) ≤ a k = k(a)
a
due to k(s) is increasing on [1, +∞) so (1) is also true.
Finally, it remains to consider the case x > a1 > 1. Consider the following
function on ( a1 , +∞):

h(x) = k(ax) + k(a) − a k(x).
We want to show that h(x) is increasing by showing that h0 (x) > 0 so h(x) ≥
h( a1 ) = 0:

h0 (x) > 0 ⇐⇒ ak 0 (ax) − a k 0 (x) > 0
2ax 2x
√ log 1+ax √ log 1+x
⇐⇒ ax > x
2k(ax) 2k(x)
⇐⇒ g(ax) > g(x)
where the function g(s) is defined on (1, +∞) by
2s
√ log 1+s
g(s) := s .
2k(s)
The final inequality g(ax) > g(x) is true because g(s) is a decreasing function
on (1, +∞) and 1 < ax < x.


You might also like