Useful Inequalities

{x2

0}

n k
k

binomial

v0.26a April 26, 2014

xi y i

Cauchy-Schwarz
i=1
n

Minkowski
i=1

|xi + yi

|xi yi |
x)r

(1 +

i=1
n

i=1

|xi

i=1

(1 +

nx)n+1

(1 +
1+

|yi

for p 1.

1/q

|yi |q

for p, q > 1,

1
p

1
q

(1 + (n +

for x [1,
1)x)n

x q
q

1+

1
),
r1

= 1.

2 x+12 x <
e

means

min{xi }

for x R, n N.
for a, b 0, n N.

for (i) x > 0, p > q > 0,

(iv) q < 0 < p , q > x > 0, (v) q < 0 < p , p < x < 0.
x n
n

ex 1 +
e x xe

1
2x

1
<

1+
xn
n!

for x R, and

ex 1 + x +
ex

1 + x,

xx

x
2

<

x2
2

x n
n

x2
n

ex 1

+ 1 ex 1 +

x n+x/2
n

x+1

2x

x y
y

yx

ex

> 1 and

> 1+

>e

x
2

for x [0, 1].

x2 1
x2 +1

x1
x

2x
2+x

ln(1 + x)

x
x+1
1
n

ln(n + 1) < ln(n) +

x2
2

ln(1 x) x
x3
2

trigonometric

hyperbolic

x cos x
2
x

x2 1
2x

ln(x)

x
2

x cos x
x3
sinh2 x

x
2

x cos x
1x2 /3

x cos2 (x/2) sin x (x cos x + 2x)/3

cosh(x) + sinh(x) ex(+x/2)

x+ y
2

for x R, [1, 1].

1
4n

(1 9n ).
n

|w i ,

x1 < 2 x2 x1

n!

1i1 <<ik n

n n 1/12n
e
e

2n
xi

1
n

wi |xi |p

1/p

for x 1.

x2
,
sinh x

n n
e

en

xi 2 max{xi }

, wi 0,

wi = 1.

for p q, wi 0.

x+y
2

x+ y
2

x+y
2

for x, y > 0.

for x, y > 0, [0, 1].

Sk

(k+1)

ai 1 ai 2 aik ,

pi (xi )

1
n

xy
ln(x)ln(y)

and

d 0.

M = mini {xi }, M = maxi {xi }.

wi |xi |q

for n d 0.

1
for (0, 2 ).

1/n

xi

2n
n

Sk+1

and

where pi 0,

for 1 k < n,

ai 0.

pi = 1, and convex.

f (ai )g(bi )pi

f (ai )pi
i=1

i=1

g(bi )pi

f (ai )g(bni+1 )pi

for a1 an , b1 bn and f, g nondecreasing, pi 0,

sin x,

x1 y +x y 1
2

p i xi

for n > k > 0.

H(x) = log2 (xx (1x)1x ).

2n

n
2

for

x+1

(xy) 4

all for x 0,

1.

wi |xi |q1

n
k

2n

Alternatively: (E [X]) E [(X)]. For concave the reverse holds.

Alternatively: E f (X)g(X) E f (X) E g(X) .

for x [0, 0.43], reverse elsewhere.

tan x

xy

i=1

x 3 cos x x x3 /6 x cos

Sk 2 Sk1 Sk+1

for x [0, 0.45], reverse elsewhere.

x3
3

wi |xi |

Jensen

<

p1

Heinz

Chebyshev

ln(n) + 1

sin x x cos(x/2) x x +

wi |xi |p

xy

Sk =
for x, y R.

n
d
i=0 i

and

i |xi

x 1, ln(x) n(x n 1) for x, n > 0.

n
1
i=1 i

x3
4

for x 0, reverse for x (1, 0].

Lehmer

for x, y > 0.

2 y exy 1 + x y + exy , and ex x + ex

n
x1
i

Maclaurin

xy
x+y

1
4

(1 8n )
n

2nH()
,
2n(1)

n n 1/(12n+1)
e
e

2n

log mean

for x, r 1.

xy

Mp Mq for p q, where Mp =

for x, n > 0.

for x (0, 1).

x1/r (x 1) rx(x1/r 1)

for G =

In the limit M0 =

for n > 1, |x| n.

for x 0, reverse for x 0.

for x [0, 1.59] and

x2

n n
e

nn
kk (nk)nk

for n1 k1 0, n2 k2 0.

d
en
for n d
d
n
d
d 1 + n2d+1
n
n n
1
i=0 i 12 n

Stirling

r > 1.

n
k

n k 0 and

square root

(ii) p < q < x < 0, (iii) q > p > x > 0. Reverse for:

exponential

en k
,
k

nd + 1

n
d
i=0 i
n
d
i=0 i
n

for x [1, 0], n N.

rx
1(r1)x

for

n
d
i=0 i

for x [0, 1], r R \ (0, 1).

(a + b)n an + nb(a + b)n1

1
p

|p

n
k

nk
k!

+n
n1 n2
n1 +k 2
k1
k1 k2
2

n
G n G
2

for x 1, r R \ (0, 1). Reverse for r [0, 1].

1
1nx

(1 + x)r 1 +

n
i=1

i=1

(1 + x)r 1 + (2r 1)x

1
p

|xi |p

1 + rx

2
yi

|p

1/p

i=1

Bernoulli

1
p

|p

Hlder
o

x2
i

i=1

n
4k!

n
k

rearrangement
i=1

ai b i

i=1

ai b(i)

ai bni+1
i=1

for a1 an ,

b1 bn and a permutation of [n]. More generally:

i=1

fi (bi )

i=1

fi (b(i) )

fi (bni+1 )
i=1

with fi+1 (x) fi (x) nondecreasing for all 1 i < n.

pi = 1.

Weierstrass

wi

1 xi

w i xi

where xi 1, and

n
k=1

Carleman

either wi 1 (for all i) or wi 0 (for all i).

Young

1
( pxp

Kantorovich

Cauchy
Hermite

U
L1

xi
yi

0<m
sum-integral

1 1
)
qxq

xi

Chong
i=1

Gibbs

i
i
n

Shapiro
i=1

yi

ai log

ai
bi

bi
ai

i=1
a
b

1
q

= 1.

f (x) dx

for f nondecreasing.

ai a i

for ai > 0.

ai , b :=

Hadamard

i=1 j=1

bi .

Hilbert

where xi > 0, (xn+1 , xn+2 ) := (x1 , x2 ),

Hardy

1
n!

Schur

Mathieu

A2
ij

n
i,j=1

2
i

where A is an n n matrix.

A2
ij

k
i=1

and

di

k
i=1

Aczl
e

n
i=1

xi i

n
ai
i=1 (1 xi )
a1 b 1

given that

n
i=2

a2
1

xi + yi

Mahler

a i xi
for xi [0, 1 ], ai [0, 1],
n
2
i=1 ai (1 xi )
2

ai b i
n
i=2

>
1/n

i=1

a2
1

a2
i
n

n=1
1
c2 +1/2

Milne

b1 min
k

i=1

or

1/n
xi

b2
1

>

n
2 b2
1
i=2 ai
n
2
i=2 bi .

+
i=1

1/n
yi

i=1

n
i=1 (ai

+ bi )

i=1

(bi )

j=1 i=1

ai(j)

a1+y b1y
i
i

1+y
a1y bi
i

for a1 a2 an and b1 bn ,

1
n!

t
i=1

ai

t
i=1 bi

for all 1 t n,

xb1 xbn
(n)
(1)

m=1

n
2
i=2 bi

ai bi b1 max
k

XA

a2
m

1
2

ai bi
n
i=1 ai +bi

n
i=1

for b1 bn 0.
ai

n=1

ak
k

n 1
|X|

pp

n=1

a2
n

2n
n=1 (n2 +c2 )2

<

p
p1

1
c2

<

an p

n=1

2 2
n=1 bn

{bk },

for am , bn R.

i=0

n
i=1 bi

Ai

n
i=1

for an 0, p > 1.

for an R.

for c = 0.

for an 0, p > 1, reverse if p (0, 1).

1, A 2[n] , no set in A is subset of another set in A.

Pr

ap
n

for c(i) depth of leaf i of binary tree, sum over all leaves.

|A| |str(A)|

Pr

n=1

n 2 a2
n

n
i

for A 2[n] , and

str(A) = {X [n] : X shattered by A},

ai

vc(A)

Sauer-Shelah

where xi , yi > 0.

i=1

an

kn

LYM
ai = 1.

a1 +a2 ++an
n

2c(i) 1

Kraft

ai

n
i=1

am bn

n=1 m+n

i=1

Abel

a1x b1+x
i
i

xa 1 xa n
(n)
(1)

n=1

n=1

for 1 k n.

A is an n n matrix. For the second inequality A is symmetric.

j=1 i=1

{bi } (majorization), i.e.

m=1

1 n the eigenvalues, d1 dn the diagonal elements.

(ai )

Carlson

Copson
n
i=1

aij

where a1 a2 an and b1 b2 bn and {ak }

where x, y, z 0, t > 0
(det A)2

j=1 i=1

and

With max{m, n} instead of m + n, we have 4 instead of .

n
2

ai(j)

xi 0 and the sums extend over all permutations of [n].

xt (x y)(x z) + y t (y z)(y x) + z t (z x)(z y) 0

for |ai |, |bi | 1.

with equality for t = n and is convex (for concave the reverse holds).

i=1

for concave, and a :=

|ai bi |

a1+x b1x
i
i

n
i=1

Karamata

for convex.

ai a(i)

n
i=1

|ak |

for 1 x y 0, and i = 1, . . . , n.

Muirhead

aij

n
k=1

where 0 ai1 aim for i = 1, . . . , n and is a permutation of [n].

and n 12 if even, n 23 if odd.

and {ai }
(a)+(b)
2

n
i=1 bi

j=1 i=1

1/k

|ai |

ai

sum-product

for ai , bi 0, or more generally:

b
a

1
p

where a < b, and convex.

(x) dx

and

xi
xi+1 +xi+2

U +1
L

f (i)

(b)

a log

for x, y 0, p, q > 0,
xi yi

n
i=1

product di.

Callebaut

U
i=L

b
a

y
q

ai
a(i)

ai

x
p

for xi , yi > 0,

A = (m + M )/2, G = mM .

M < ,

1
ba

A
G

f (b)f (a)
ba

a+b
2

xy

f (x) dx

(a)

wi 1 and xi 1, the reverse holds.

k
i=1

(1)j1 Sj

for 1 k n, k odd,

(1)j1 Sj

for 2 k n, k even.

j=1
k

Ai

j=1

Sk =
1i1 <<ik n

vc(A) = max{|X| : X str(A)}.

Pr Ai1 Aik

where Ai are events.

Bhatia-Davis

Var[X] M E[X] E[X] m

where X [m, M ].

Samuelson

n 1 xi + n 1

for i = 1, . . . , n.

xi /n , 2 =

Where =
Markov

Pr |X| a E |X| /a

Pr X E[X] t

VysochanskijPetunin-Gauss

for X [0, 1] and c 0, E[X] .

2nd

moment

Pr X > 0

(E[X])2 /(E[X 2 ])

kth

moment

Pr X t

Cherno

Pr[X t] F (a)/at

F (z) =

pk

zk

pk

exp

2
2

(1+)
pk /

k k = /(1 p), E[Xi ] = pi , X =

Azuma

Pr X E[X] 2 exp
Xi [ai , bi ] (w. prob. 1), X =
E eX exp

Pr max |Sk |
k

for Xi [0, 1] k-wise i.r.v.,

,
n

2
2(n 2 + M /3)

exp

for Xi i.r.v.,

Pr Xn X0 2 exp

n
2
i=1 ci

1
n

Var[Xi ], 0.

for martingale (Xk ) s.t.

Var[Z]

1
2

E
i=1

Z Z (i)

for Xi , Xi X i.r.v.,

Pr Z E[Z] 2 exp

2 2
n
2
i=1 ci

for Xi , Xi X i.r.v.,

Z, Z (i) as before, s.t. Z Z (i) ci for all i, and 0.

> 0.
Janson

for Xi i.r.v.,

M Pr
M =

Xi , 0.

Lovsz
a

a)2

Pr

B i M exp

(1 Pr[Bi ]), =
Bi

2 2

i=j,Bi Bj

(1 xi ) > 0

where Pr[Bi ] for all i,

where Pr[Bi ] xi

(i,j)D

(1 xj ),

for xi [0, 1) for all i = 1, . . . , n and D the dependency graph.

1
Var[Sn ]
2

1
2

If each Bi mutually indep. of the set of all other events, exc. at most d,

Var[Xi ]

Pr[Bi ] p for all i = 1, . . . , n , then if ep(d + 1) 1 then Pr

where X1 , . . . , Xn are i.r.v., E[Xi ] = 0,

Xi

f : X n R, Z = f (X1 , . . . , Xn ), Z (i) = f (X1 , . . . , Xi , . . . , Xn ).

Xi .

A related lemma, assuming E[X] = 0, X [a, b] (w. prob. 1) and R:

Pr

Xi Xi1 < ci (w. prob. 1), for i = 1, . . . , n, 0.

for R 2e ( 5.44).

2 2
n
2
i=1 (bi ai )

Var[Xi ], |Xi | M (w. prob. 1), 0,

E[Xi ] = 0, |Xi | < M (w. prob. 1) for all i, 2 =

for [0, 1).

Xi , = E[X], p =

where Xi i.r.v.,

(u) = (1 + u) log(1 + u) u.
i=1

Xi , = E[X], 0.

1
n

for martingale (Xk ) and > 0.

n 2
M

M2
n 2

exp

E[Xi ] = 0, 2 =

McDiarmid

Pr X (1 + )

Kolmogorov

i=1

Xi

for Xi {0, 1} k-wise i.r.v., E[Xi ] = p, X =

0.

Pr

Bernstein

2
min{2 + , 3}

exp

2
.
3

Pr max1kn |Xk | E |Xn | /

for X r.v., Pr[X = k] = pk ,

e
(1 )(1)

t
k

1kn
k
i=1 Xi ,

Bennett

Simpler (weaker) form: Pr X R 2R

1kn

Doob

k/2

where Xi indep. r.v. drawn from {0,1}, X =

Pr X t

if

Pr max |Sk | 3 3 max Pr |Sk |

and

e
(1 + )(1+)

Pr X (1 )

2
,
3

if

where Xi are i.r.v., Sk =

probability gen. func., and a 1.

Pr X (1 + )

Hoeding

Etemadi

for Xi [0, 1] k-wise indep. r.v.,

Xi , i = 1, . . . , n, = E[X], Ck = 2 ke1/6k 1.0004, k even.

Pr X t Ck
X=

where t > 0.

where E[X 2 ] = 0.

E (X )k
tk
nk
t2

4 2
92

8
,
3

if

2 = Var[X] < , 2 = Var[X] + (E[X] m)2 = E[(X m)2 ].

where E[X] 0.

Pr X = 0 Var[X]/(E[X 2 ])

for X 0,

Where X is a unimodal r.v. with mode m,

where t > 0.

Pr X E[X] t Var[X]/(Var[X] + t2 )

4
92

Pr X E[X]

Pr X m 1

for f 0, and f (x) s > 0 for all x S.

Var[X]/t2

Var[X]
(1 )2 (E[X])2 + Var[X]

Pr X m

where X is a random variable (r.v.), a > 0.

Pr X S] E[f (X)]/s

Pr X E[X]

Var[X] < , and (0, 1).

(xi )2 /n.

Pr X c (1 E[X])/(1 c)

Chebyshev

Paley-Zygmund

k
i=1

Xi and > 0.

Lszl Kozma latest version: http://www.Lkozma.net/inequalities_cheat_sheet

B i > 0.