Professional Documents
Culture Documents
Smoothing Handout PDF
Smoothing Handout PDF
Ricky
September 2019
§1 Introduction
In most inequality problems we are asked to find the minimum (or maximum) of a
function f (a1 , a2 , · · · , an ), where a1 , a2 , · · · , an is a set of real numbers. Let b1 , b2 , · · · , bn
be the values of a1 , a2 , · · · , an respectively when f obtains its minimum, then what we
want to prove is equivalent to f (a1 , a2 , · · · , an ) ≥ f (b1 , b2 , · · · , bn ). We are therefore
motivated to find an operation P such that
• The inequality f (a1 , a2 , · · · , an ) ≥ f (P (a1 , a2 , · · · , an )) easily holds,
• P (a1 , a2 , · · · , an ) preserves any condition regarding a1 , a2 , · · · , an ,
• There exists an integer k ≥ 2 for which P (k) (a1 , a2 , · · · , an ) = (b1 , b2 , · · · , bn ).
(This doesn’t have to be true; Sometimes P (k) (a1 , a2 , · · · , an ) = (b1 , b2 , · · · , bn−2 , t1 , t2 )
also works because then we only have to prove a 2-variable inequality.)
Finding the operation P and applying it to a1 , a2 , · · · , an so that it ultimately becomes
b1 , b2 , · · · , bn is known as “smoothing”. It’s the most useful technique when dealing with
n-variable inequalities that are somewhat symmetric.
§2 Elementary Applications
Why is this? Let us discover the advantage of this method through the following simple
example:
Example 2.1
Let x, y, z ∈ [1, 2], prove that
( )
1 1 1 18 1 1 1
+ + + ≥6 + + .
x y z x+y+z x+y y+z z+x
1
Ricky (September 2019) Smoothing
In order to make the above expression always larger than zero, let x = max{x, y, z},
then (t + x)(x + y)(x + z) > 6yzt, and consequently f (x, y, z) ≥ f (x, t, t). Now we have
transformed the original 3-variable inequality into a two-variable one:
( )
1 2 18 2 1
+ + ≥6 + .
x t x + 2t x + t 2t
The proof of this, as can be sensed, is very simple.
This above example illustrated some very amazing things that can be done by smoothing.
First, by taking an operation P that only alters the value of two variables each time, the
inequality f (x, y, z) ≥ f (P (x, y, z)) is usually easy to prove because many of the terms
cancel out, and the remaining terms usually have factors like (y−z)2 thanks to symmetry.
Then, we can transform a multi-variable inequality into a two- or one-variable inequality.
Second, in the above example f (x, y, z) ≥ f (P (x, y, z)) does not necessarily hold for
any permutation of x, y, z. However, there are various ways to fix this. One, for instance,
is to create extra conditions such as x = max{x, y, z}.
Third, finding
( the operation ) P doesn’t require much thought. In many inequalities
y+z y+z √ √
the classical x, , , (x, yz, yz), (x, y + z, 0) are most useful.
2 2
Hence many people find “smoothing” to be a very useful way to prove inequalities
because it doesn’t require much thought. Proofs with AM-GM and Cauchy need ob-
servation skills, but with enough time, experience and courage, anyone can prove an
inequality through smoothing. Now let’s see some other mechanic examples.
Proof. Apparently the equality occurs when two of x, y, z is 1 and the other is 0. Let
1 1 1
f (x, y, z) = + + , then in order to change one variable to 0, we consider
x(+ y y + z z)+ x
1 1 − xy
P (x, y, z) = x + y, , 0 . Using z = we have
x+y x+y
( ) [ ]
1 2 + x2 + y 2 1
f (x, y, z) − f x + y, , 0 = (x + y) −1−
x+y (1 + x2 )(1 + y 2 ) 1 + (x − y)2
2 − 2xy − (x + y)2 xy
= xy(x + y) ·
(1 + x2 )(1 + y 2 )(1 + (x + y)2 )
2z − (x + y)xy
= xy(x + y)2 · .
(1 + x2 )(1 + y 2 )(1 + (x + y)2 )
Again, to make the expression
( ) let z = max{x, y, z}, then 2z ≥ x+y ≥ (x+y)xy,
positive,
1
hence f (x, y, z) ≥ f x + y, , 0 . It remains to prove that
x+y
( )
1 1 1 5
f x + y, ,0 = x + y + + ≥ ,
x+y x+y 1 2
x+y+
x+y
which is trivial.
2
Ricky (September 2019) Smoothing
Proof. This example is a bit more technical. We see that the equality is obtained when
√ √
a = b = c = d = 1, so consider P (a, b, c, d) = ( ac, b, ac, d) with f (a, b, c, d) = LHS.
Then we have
√ √ 1 1 9 2 9
f (a, b, c, d) − f ( ac, b, ac, d) = + + −√ − √
a c a+b+c+d ac 2 ac + b + d
√ √ 2 √
( a − c) ((a + b + c + d)(2 ac + b + d) − 9ac)
= √
ac(a + b + c + d)(2 ac + b + d)
Exercise 2.5 (2011 CMO). Let n ≥ 4 be an integer. For non-negative real numbers
a1 , a2 , · · · , an , b1 , b2 , · · · , bn such that a1 + a2 + · · · + an = b1 + b2 + · · · + bn , find the
maximum of
∑n
ai (ai + bi )
i=1
.
∑
n
bi (ai + bi )
i=1
3
Ricky (September 2019) Smoothing
Example 3.1
For positive reals a, b, c, prove that
63 (a + b + c)(a2 + b2 + c2 ) 27 a + b + c
+ ≥ · √3
.
2 abc 2 abc
Proof. Denote f (a, b, c) = LHS − RHS, we will leave to the reader as an exercise to show
that √ √
f (a, b, c) ≥ f (a, bc, bc)
holds for all a, b, c. What we’ve done in the above examples is to directly expand and
bash after this. But now let us introduce a more advanced approach based on the
following powerful:
f (x1 , x2 , x3 , · · · , xn ) ≥ f (A, A, A, · · · , A)
4
Ricky (September 2019) Smoothing
f (x1 , x2 , x3 , · · · , xn ) ≥ f (G, G, G, · · · , G)
§3.2 Choice of P
) (
a+b a+b √ √
Not every problem can be easily dealt with using the , or the ( ab, ab)
2 2
trick. Sometimes plugging in actual values could massively reduce the amount of com-
putation needed. This method is also extremely advantageous as it could approach the
equality case easily and save us the need to prove theorem 3.2.
Example 3.5
Let a, b, c, d be non-negative reals such that a + b + c + d = 4. Prove that
1
bcd + cda + dab + abc − abcd ≤ (ab + ac + ad + bc + bd + cd).
2
5
Ricky (September 2019) Smoothing
Note that algebraic manipulation and the choice of P are both crucially important.
Exercise 3.6. Let x1 , x2 , x3 , x4 be non-negative reals such that x1 + x2 + x3 + x4 = 1.
Find the range of
∑4 ∑
S= x3i − 6 xi xj xk .
i=1 1≤i<j<k≤4
§3.3 Casework
The idea of this section can be best understood through the following example:
Example 3.7
Let real numbers x1 , x2 , x3 , x4 , y1 , y2 satisfy
y2 ≥ y1 ≥ x4 ≥ x3 ≥ x2 ≥ x1 ≥ 2, x1 + x2 + x3 + x4 ≥ y1 + y2 .
Prove that x1 x2 x3 x4 ≥ y1 y2 .
• This problems concerns solely the product of four numbers that have a fixed sum.
Given any two numbers with a fixed sum, their product could be easily determined.
So taking operations like P (x1 , x2 , x3 , x4 ) = (a, x1 + x2 − a, x3 , x4 ) would be very
useful. Here a = y1 and a = 2 could both work.
Now let us start with the problem. The first idea we came up with is P (x1 , x2 , x3 , x4 ) =
(y1 , x1 + x2 − y1 , x3 , x4 ); this is feasible only if x1 + x2 − y1 ≥ 2. If so, x1 x2 x3 x4 ≥
y1 x2 x3 (x1 + x4 − y1 ) > y1 (x1 + x2 + x3 + x4 − y1 ) = y1 y2 , and we are done.
If x1 + x2 − y1 < 2, let’s consider P (x1 , x2 , x3 , x4 ) = (2, x2 , x3 , x4 + x1 − 2). We have
6
Ricky (September 2019) Smoothing
16
If 4y1 (y2 − 4) ≥ y1 y2 , i.e. y2 ≤
, then we are done again.
3
16
Now the only remaining case is 4 < y2 < , which seems easy. We use the rough
3
estimate that
(y1 + y2 )4
x1 x2 x3 x4 ≥ 8(x1 + x2 + x3 + x4 − 6) ≥ 8(y1 + y2 − 6) ≥ ≥ y1 y2 .
4
What’s remarkable about this problem is that we didn’t really use smoothing any-
where; We used merely the identity that xy < K(x + y − K) if x < K < y. As we will
see in some future examples, this identity is extremely important.
§3.4 Be careful
In this section I will show you a smoothing approach that turns out to be... a fake-
solve.
Example 3.9
For a given positive integer n, find the greatest constant kn such that
Proof. Obviously kn = 27n2 is the optimal constant, with equality obtained when
a1 = a2 = · · · = a)3n . One option would be to use the operation P (ai , ai+1 ) =
(
ai + ai+1 ai + ai+1
, . But with the annoying condition that a1 ≤ a2 ≤ · · · ≤ a3n , we
2 2
can only operate on adjacent variables ai , ai+1 , which is very inconvenient. So we “clev-
erly” consider the operation P (ai , ai+1 ) = (ai−1 , ai+1 +ai −ai−1 ) or (ai+1 +ai −ai+2 , ai+2 ).
To see this visually, consider ai−1 , ai , ai+1 , ai+2 on the number axis. The numbers ai and
ai+1 are moving towards ai−1 and ai+2 , respectively, at the same speed. One of them
would first reach their destination; When this happens the two numbers stop moving.
Then the order condition still holds, and two numbers in the sequence {ai } have become
equal.
The rest should be long but easy. Right?
This is incorrect for various reasons: Firstly, the expression is not symmetric for any
two variables, so sometimes this operation doesn’t work. Secondly, at some point it may
be impossible to further operate. For instance if you get a1 = a2 = a3 < a4 = a5 = a6 <
· · · < a3n−2 = a3n−1 = a3n , there’s nothing you can do. The operation I provided in this
section is very dangerous indeed; It’s often the beginning of a false proof.
Exercise 3.11. Find an alternative solution to Example 3.6 using the following fact:
For 0 ≤ x ≤ K ≤ y ≤ L ≤ z, xyz ≤ KL(x + y + z − K − L).
7
Ricky (September 2019) Smoothing
∑ reals x1 , x2 , · · · , xn such
Let f (x) = (x + a)(x + b) where a, b > 0. For non-negative
that x1 + x2 + · · · + xn = 1, find the maximum of F = 1≤i<j≤n min{f (xi ), f (xj )}.
Proof. You must have gotten pretty bored by now; This is just another example where
the equality case is x1 = x2 = · · · = xn . WLOG let x1 ≤ x2 ≤ · · · ≤ xn , then we can
write
∑n
F = (n − i)f (xi ).
i=1
( )
xi + xj
Hence it suffices to prove (2n − i − j)f ≥ (n − i)f (xi ) + (n − j)f (xj ). But
2
this isn’t necessarily true!
Since the most straightforward approach failed, let’s try induction. We will prove
inductively that regardless of x1 + x2 + · · · + xk , the maximum of Fk (x1 , x2 , · · · , xk ) is
obtained when x1 = x2 = · · · = xk =. When k = 2 this is obviously true. Suppose this
is true for k, WLOG let x1 ≤ x2 ≤ · · · ≤ xk+1 , then f (x1 ) ≤ f (x2 ) ≤ · · · ≤ f (xk+1 ).
Thus
Fk+1 (x1 , x2 , · · · , xk+1 ) = kf (x1 ) + Fk (x2 , x3 , · · · , xk+1 ).
According to our inductive hypothesis, this expression is maximal when x2 = x3 = · · · =
S − x1
xk+1 = . Hence
k
k(k − 1) S − x1
Fk+1 (x1 , x2 , · · · , xk+1 ) ≤ kf (x1 ) + f( ),
2 k
S
which is a quadratic expression regarding x1 . Since 0 ≤ x1 ≤ , and the leading
k+1
coefficient is positive, the maximum of this quadratic is obtained when x1 = 0 or x1 =
S S
. Some simple computation yields that x1 = , and thus our statement is true
k+1 k+1
for k + 1. ( )
n(n − 1) 1 n−1
Now we have proven max F = f = (na + 1)(nb + 1).
2 n 2n
Exercise 3.13. Use Jensen’s Inequality to solve Example 3.11.
Exercise 3.14 (2017 CMO). Given an integer n ≥ 2 and real numbers a, b such that
0 < a < b. Let x1 , x2 , . . . , xn ∈ [a, b] be real numbers. Find the maximum value of
x21 x22 x2n−1 x2n
x2 + x3 + ··· + xn + x1
.
x1 + x2 + · · · + xn−1 + xn
§4 Selected Problems
In the following collection of exercises, the first few problems are warm-ups; but you
should pay some attention to your proof-writing. Then there are a few applications of
Generalized Jensen’s, a technique I found unnecessary to elaborate on. The last few
problems are the ones that appeared in actual tests.
8
Ricky (September 2019) Smoothing
Exercise 4.4 (1989 CMO). Given 1989 points in the space, any three of which are not
collinear. We divide these points into 30 groups such that the numbers of points in these
groups are different from each other. Consider triangles whose three vertices are in three
different groups among the 30. Determine the numbers of points in each group such that
the number of such triangles attains its maximum.
Exercise 4.5. Let x, y, z, w be positive reals such that x + y + z + w = 4. Prove that
1 1 1 1 13 21
+ + + +√ ≥ .
x y z w 2 2 2
x +y +z +w 2 2
3 + a2 + b2 + c2 ≥ 2(ab + bc + ca).
Exercise 4.11 (2017 China First Round). Let x1 , x2 , x3 be non-negative reals such that
x1 + x2 + x3 = 1. Find the minimum and maximum of
( x2 x3 )
(x1 + 3x2 + 5x3 ) x1 + + .
3 5
9
Ricky (September 2019) Smoothing
Exercise 4.12 (2019 China Second Round). Let a1 , a2 , · · · , an be integers such that
1 = a1 ≤ a2 ≤ · · · ≤ a2019 = 99. Find the minimum f0 of the expression
a1 ≥ a2 ≥ · · · ≥ am ≥ b1 ≥ b2 ≥ · · · ≥ bl ≥ 0,
a1 + a2 + · · · + am ≤ n, b1 + b2 + · · · + bl ≤ n.
Find the maximum of a21 + a22 + · · · + a2m + b21 + b22 + · · · + b2l .
(Turn the condition into xi xj = 4−|i−j| for some i, j. Prove that in the above equation,
|i − j| can obtain 1, and then find a bound for every xi so that there’s only one variable
left. Alternatively, use induction.)
10