the usual form of Jensen’s inequality
For a given convex function
f
f
f, the below inequality is satisfied:
f
(
t
x
1
+
(
1
−
t
)
x
2
)
≤
t
f
(
x
1
)
+
(
1
−
t
)
f
(
x
2
)
f(tx_1+(1-t)x_2)\le tf(x_1)+(1-t)f(x_2)
f(tx1+(1−t)x2)≤tf(x1)+(1−t)f(x2)
where
t
t
t satisfies
t
∈
[
0
,
1
]
t\in [0,1]
t∈[0,1].
(Note: the inequality is reversed if
f
f
f is concave,
f
(
t
x
1
+
(
1
−
t
)
x
2
)
≥
t
f
(
x
1
)
+
(
1
−
t
)
f
(
x
2
)
f(tx_1+(1-t)x_2)\ge tf(x_1)+(1-t)f(x_2)
f(tx1+(1−t)x2)≥tf(x1)+(1−t)f(x2))
Proof1
We might as well define
F
(
x
)
F(x)
F(x) as:
F
(
x
)
=
f
(
t
x
1
+
(
1
−
t
)
x
)
−
t
f
(
x
1
)
−
(
1
−
t
)
f
(
x
)
F(x)=f(tx_1+(1-t)x)- tf(x_1)-(1-t)f(x)
F(x)=f(tx1+(1−t)x)−tf(x1)−(1−t)f(x)
where x 1 x_1 x1 is a real number and x ≥ x 1 x \ge x_1 x≥x1.
Then take the derivative of
F
(
x
)
F(x)
F(x) with respect to
x
x
x :
F
′
(
x
)
=
(
1
−
t
)
f
′
(
t
x
1
+
(
1
−
t
)
x
)
−
(
1
−
t
)
f
′
(
x
)
=
(
1
−
t
)
(
f
′
(
x
+
t
(
x
1
−
x
)
)
−
f
′
(
x
)
)
\begin{matrix} F'(x)&=& (1-t)f'(tx_1+(1-t)x)-(1-t)f'(x) \\ \\ &=& (1-t)(f'(x+t(x_1-x))-f'(x)) \end{matrix}
F′(x)==(1−t)f′(tx1+(1−t)x)−(1−t)f′(x)(1−t)(f′(x+t(x1−x))−f′(x))
where
x
+
t
(
x
1
−
x
)
≤
x
x+t(x_1-x) \le x
x+t(x1−x)≤x.
According to the properties of convex functions, it can be seen that
f
′
(
x
+
t
(
x
1
−
x
)
)
≤
f
′
(
x
)
f'(x+t(x_1-x))\le f'(x)
f′(x+t(x1−x))≤f′(x) is satisfied.
So
F
′
(
x
)
≤
0
F'(x)\le 0
F′(x)≤0,then:
F
(
x
)
≤
F
(
x
1
)
=
f
(
t
x
1
+
(
1
−
t
)
x
1
)
−
t
f
(
x
1
)
−
(
1
−
t
)
f
(
x
1
)
=
f
(
x
1
)
−
f
(
x
1
)
=
0
\begin{matrix} F(x)\le F(x_1)&=&f(tx_1+(1-t)x_1)- tf(x_1)-(1-t)f(x_1) \\ \\ &=&f(x_1)-f(x_1) \\ \\ &=&0 \end{matrix}
F(x)≤F(x1)===f(tx1+(1−t)x1)−tf(x1)−(1−t)f(x1)f(x1)−f(x1)0
F
(
x
)
≤
0
f
(
t
x
1
+
(
1
−
t
)
x
)
−
t
f
(
x
1
)
−
(
1
−
t
)
f
(
x
)
≤
0
f
(
t
x
1
+
(
1
−
t
)
x
)
≤
t
f
(
x
1
)
−
(
1
−
t
)
f
(
x
)
\begin{matrix} F(x)&\le& 0 \\ \\ f(tx_1+(1-t)x)- tf(x_1)-(1-t)f(x)&\le& 0 \\ \\ f(tx_1+(1-t)x)&\le& tf(x_1)-(1-t)f(x) \end{matrix}
F(x)f(tx1+(1−t)x)−tf(x1)−(1−t)f(x)f(tx1+(1−t)x)≤≤≤00tf(x1)−(1−t)f(x)
the generalized form of Jensen’s inequality
f ( t 1 x 1 + t 2 x 2 + . . . + t n x n ) ≤ t 1 f ( x 1 ) + t 2 f ( x 2 ) + . . . + t n f ( x n ) f(t_1x_1+t_2x_2+...+t_nx_n)\le t_1f(x_1)+t_2f(x_2)+...+t_nf(x_n) f(t1x1+t2x2+...+tnxn)≤t1f(x1)+t2f(x2)+...+tnf(xn)
where t 1 , t 2 . . . t n t_1,t_2...t_n t1,t2...tn are nonnegative real number such that ∑ i n t i = 1 \sum_i^nt_i=1 ∑inti=1.
When n = 2 n=2 n=2, the generalized form become to the usual form.
Proof2
Suppose the equation holds when n = k n=k n=k, that is, f ( t 1 x 1 + t 2 x 2 + . . . + t k x k ) ≤ t 1 f ( x 1 ) + t 2 f ( x 2 ) + . . . + t k f ( x k ) f(t_1x_1+t_2x_2+...+t_kx_k)\le t_1f(x_1)+t_2f(x_2)+...+t_kf(x_k) f(t1x1+t2x2+...+tkxk)≤t1f(x1)+t2f(x2)+...+tkf(xk).
Then when
n
=
k
+
1
n=k+1
n=k+1,
t
1
f
(
x
1
)
+
t
2
f
(
x
2
)
+
.
.
.
+
t
k
+
1
f
(
x
k
+
1
)
=
(
1
−
t
k
+
1
)
{
1
(
1
−
t
k
+
1
)
[
t
1
f
(
x
1
)
+
t
2
f
(
x
2
)
+
.
.
.
+
t
k
f
(
x
k
)
]
}
+
t
k
+
1
f
(
x
k
+
1
)
=
(
1
−
t
k
+
1
)
[
t
1
(
1
−
t
k
+
1
)
f
(
x
1
)
+
t
2
(
1
−
t
k
+
1
)
f
(
x
2
)
+
.
.
.
+
t
k
(
1
−
t
k
+
1
)
f
(
x
k
)
]
+
t
k
+
1
f
(
x
k
+
1
)
≥
(
1
−
t
k
+
1
)
f
(
t
1
(
1
−
t
k
+
1
)
x
1
+
t
2
(
1
−
t
k
+
1
)
x
2
+
.
.
.
+
t
k
(
1
−
t
k
+
1
)
x
k
)
+
t
k
+
1
f
(
x
k
+
1
)
\begin{matrix} t_1f(x_1)+t_2f(x_2)+...+t_{k+1}f(x_{k+1}) &=&(1-t_{k+1})\{\frac{1}{(1-t_{k+1})}[t_1f(x_1)+t_2f(x_2)+...+t_kf(x_k)]\}+t_{k+1}f(x_{k+1}) \\ \\ &=&(1-t_{k+1})[\frac{t_1}{(1-t_{k+1})}f(x_1)+\frac{t_2}{(1-t_{k+1})}f(x_2)+...+\frac{t_k}{(1-t_{k+1})}f(x_k)]+t_{k+1}f(x_{k+1}) \\ \\ &\ge &(1-t_{k+1})f(\frac{t_1}{(1-t_{k+1})}x_1+\frac{t_2}{(1-t_{k+1})}x_2+...+\frac{t_k}{(1-t_{k+1})}x_k)+t_{k+1}f(x_{k+1}) \\ \\ \end{matrix}
t1f(x1)+t2f(x2)+...+tk+1f(xk+1)==≥(1−tk+1){(1−tk+1)1[t1f(x1)+t2f(x2)+...+tkf(xk)]}+tk+1f(xk+1)(1−tk+1)[(1−tk+1)t1f(x1)+(1−tk+1)t2f(x2)+...+(1−tk+1)tkf(xk)]+tk+1f(xk+1)(1−tk+1)f((1−tk+1)t1x1+(1−tk+1)t2x2+...+(1−tk+1)tkxk)+tk+1f(xk+1)
We might as well define u 1 = t 1 ( 1 − t k + 1 ) x 1 + t 2 ( 1 − t k + 1 ) x 2 + . . . + t k ( 1 − t k + 1 ) x k u_1=\frac{t_1}{(1-t_{k+1})}x_1+\frac{t_2}{(1-t_{k+1})}x_2+...+\frac{t_k}{(1-t_{k+1})}x_k u1=(1−tk+1)t1x1+(1−tk+1)t2x2+...+(1−tk+1)tkxk and u 2 = x k + 1 u_2=x_{k+1} u2=xk+1.
From the Proof1, we can get f ( ( 1 − t k + 1 ) u 1 + t x + 1 u 2 ) ≤ ( 1 − t k + 1 ) f ( u 1 ) + t k + 1 f ( u 2 ) f((1-t_{k+1})u_1+t_{x+1}u_2)\le (1-t_{k+1})f(u_1)+t_{k+1}f(u_2) f((1−tk+1)u1+tx+1u2)≤(1−tk+1)f(u1)+tk+1f(u2).
t 1 f ( x 1 ) + t 2 f ( x 2 ) + . . . + t k + 1 f ( x k + 1 ) ≥ ( 1 − t k + 1 ) f ( u 1 ) + t k + 1 f ( u 2 ) ≥ f ( ( 1 − t k + 1 ) u 1 + t x + 1 u 2 ) = f ( t 1 x 1 + t 2 x 2 + . . . + t n x n ) t_1f(x_1)+t_2f(x_2)+...+t_{k+1}f(x_{k+1})\ge(1-t_{k+1})f(u_1)+t_{k+1}f(u_2) \ge f((1-t_{k+1})u_1+t_{x+1}u_2)=f(t_1x_1+t_2x_2+...+t_nx_n) t1f(x1)+t2f(x2)+...+tk+1f(xk+1)≥(1−tk+1)f(u1)+tk+1f(u2)≥f((1−tk+1)u1+tx+1u2)=f(t1x1+t2x2+...+tnxn)
We deduce the inequality is true for n + 1 n+1 n+1, and we have proved the correctness for n = 2 n=2 n=2, by the principal of mathematical induction it follows that the result is also true for all integer n n n greater than 2.