Useful Inequalities

n 2 n

{x2
n

0}
2 yi

v0.20b July 22, 2013

square root Stirling

2 x + 12 x < e
n n e x3 2

1 x

<

x + 1

x 1 < 2 x2 x 1 2n
n n 1/12n e e

for x 1.
n n e

Cauchy-Schwarz
i=1 n

x i yi

x2 i
i=1
1 p

i=1 n
1 p

2n

n n 1/(12n+1) e e x cos x 1x2 /3

n!

en
x 3

1 p

Minkowski
i=1 n

|xi + y i |p
n

i=1

| xi | p
n i=1

+
i=1 1/q

|y i | p

trigonometric for p 1.
1 p 1 q

x
2 x

x cos x
x sinh2 x
3

x 3 cos x x x3 /6 x cos
x3 3

sin x,
2

hyperbolic = 1. Napier

x cos x

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

x2 , sinh x

1/p

H older
i=1

|x i y i | x)r

i=1

| xi |

|y i |

for p, q > 1,

sin x x cos(x/2) x x +
ba ln(b)ln(a)

tan x

all for x 0,

cosh(x) + sinh(x) a< ab <

for x R, [1, 1]. where 0 < a < b. where x, y > 0, [0, 1].
1 n

Bernoulli

(1 + (1 +

x ) 2n x)n

1 + rx 1 + 2nx
1 1nx

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

<b

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

Heinz means

xy <

x 1 y +x y 1 2 n 1 x i 1/p

<

x +y 2

for x [1, 0], n N. r > 1. min{xi }

(
i

xi )1/n
1/q

xi

1 n

xi 2 max{xi } wi = 1.

(1 + nx)n+1 (1 + (n + 1)x)n exponential ex 1 + e x xe e x 1 + x, and

for x R, n N. ex 1
x2 n x n+x/2 n

power means for n > 1, |x| n. for x, n > 0. Maclaurin

1+

w i xp i

w i xq i

for xi 0, p q, and
(k+1)

for x R,
x 2 x 2

ex 1 +

ex 1 + x +
1 2 x

< xx < x2 x + 1,

for x [0, 1.59] and 2x 1 for x (0, 1). for x, r 1. for x, y R.

for x 0, reverse for x 0.

Sk 2 Sk1 Sk+1 1 Sk = n
k

and

Sk

Sk+1

for 1 k < n,

x 2

for x [0, 1]. Jensen

1i1 <<ik n

ai 1 ai 2 ai k ,

and

ai 0. pi = 1, and convex.

x1/r (x 1) rx(x1/r 1) xy + y x > 1 for x, y > 0.

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

p i xi

p i ( xi )

where pi 0,

2 y exy 1 + x y + exy logarithm

Chebyshev
i=1

f (a i )g (b i )p i

f (a i )p i
i=1 i=1

g (b i )p i

i=1

f (ai )g (bni+1 )pi pi = 1.

x2 1 x2 +1

ln(x)
x2 2 x3 4

x2 1 2x

x 1 for x > 0.

ln(1 + x)

x x+1

for x 0, reverse for x (1, 0]. for x [0, 0.45], reverse elsewhere. for x [0, 0.43], reverse elsewhere. +1 for x > 0, and q p > 0.
n k

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

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

ln(1 + x) x ln(1 x) ln(n + 1) ln(x) binomial

rearrangement
i=1

xp 1 p n k n n

2 3 x x x 2 2 n 1 i=1 i ln(n)

ai bi

i=1

ai b (i)

i=1

ai bni+1
n

for a1 an ,

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

xq 1 q

xq ln(x)
en k k

f i (b i )

i=1

f i (b ( i ) )

i=1

fi (bni+1 )

n2 k2

nk k!

and

< 2n

G for G =
n 1 +n 2 k 1 +k 2

n (1)(1)n , 2n(1)

for n k > 0.
1 ]. (0, 2

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

xi 2
xi yi

yi 2

for n1 k1 0, n2 k2 0. and
n d i=0 i

0<m Young Cauchy

M < ,

for xi , yi > 0, A = (m + M )/2, G = mM .

A G

xi y i

binomial sum

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

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

nd + 1

2n

for n d 0.

1 1 1 ( px xy p + qxq )

+y q

for x, y 0, p, q > 0 and

1 +1 p q

= 1.

1.

1 for (0, 2 ).

(a )

f (b ) f (a ) ba

(b )

where a < b, and convex.

Weierstrass

either wi 1 (for all i) or wi 0 (for all i). If wi [0, 1], Chong Hermite Gibbs Woeginger Pe cari c
ai n i=1 a (i) a +b 2

1 xi

wi

w i xi

where xi 1, and

Callebaut

x 1 x a1+ bi i

x 1+x a1 bi i

y 1 y a1+ bi i

y 1+y a1 bi i

wi 1 and xi 1, the reverse holds. and

for 1 x y 0, and i = 1, . . . , n. for ai > 0. Karamata

n
1 ba

ai a i

n i=1

ai a (i)

(a i )

n i=1

(b i )

for a1 a2 an and b1 bn ,
t i=1

ai bi

(x) dx
a b b a

(a)+(b) 2

for convex. ai , b := bi . Muirhead

with equality for t = n and is convex (for concave the reverse holds).
1 n!
n 1 xa xa (n ) (1)

and {ai }

{bi } (majorization), i.e.

ai

t i=1 bi

for all 1 t n,

ai log ai

a log a

for ai , bi 0, a :=

n 1 xb xb (n ) (1)

bi ai

where a1 a2 an and b1 b2 bn and {ak } xi 0 and the sums extend over all permutations of [n].
m=1 am bn n=1 m+n
1 2 1

{ b k },

for concave and variables as before. where either (i) x > 0, p > q > 0, Hilbert

1+

x p p

1+

x q q

m=1

a2 m

2 2 n=1 bn

for am , bn R.

(ii) p < q < x < 0 or (iii) q > p > x > 0. Reverse, if (iv ) q < 0 < p , q > x > 0 or (v ) q < 0 < p , p < x < 0. Shapiro
xi n i=1 xi+1 +xi+2

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

n 2

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

a2 n <

p p 1 n=1 1 c2

n=1

ap n

for an 0, p > 1.

and n 12 if even, n 23 if odd. Schur where x, y, z 0, t > 0 Hadamard Schur (det A)2
n i=1 n i=1 n i,j =1

an <

n=1

n 2 a2 n

for an R.

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

2n n=1 (n2 +c2 )2 n=1

for c = 0.

A2 ij and

where A is an n n matrix.
k i=1

Copson

2 i

A2 ij

di

k i=1

ak k

pp

an p

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

for 1 k n. Kraft

1 n the eigenvalues, d1 dn the diagonal elements. Ky Fan Acz el

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

2 c( i ) 1 n |X |
1

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

1 ], ai [0, 1], for xi [0, 2

ai = 1.

LYM
X A

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

given that a2 1 > Mahler

a1 b 1

n i=2

ai b i
n i=2 1/n

2 a2 i or b1 > n i=1

a2 1 xi

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

n 2 i=2 bi

Sauer-Shelah

|A|

vc(A) i=0

n , vc(A) is the size of the largest set shattered by A. i

xi + y i
k

1/n

n i=1 k

yi

1/n

where xi , yi > 0. Bonferroni Pr

n i=1

Ai Ai

j =1 k j =1

(1)j 1 Sj (1)j 1 Sj

for 1 k n, k odd, for 2 k n, k even. where Ai are events.

Abel

bn min
k i=1 n i=1 (ai

|a i | + bi )

i=1

|ai bi | bn max
k ai bi n i=1 ai +bi

i=1

|a i | ai

for 0 b1 bn .
n i=1 bi

Pr
i=1

Milne

|a k |

n i=1

Sk =
1i1 <<ik n

Pr Ai1 Aik

Carleman

n k=1 m n

k i=1

|a i |
m

1/k

n k=1

Bhatia-Davis
m n m n

Var[X ] M E[X ] E[X ] m n 1 xi + n 1 xi /n , 2 = (xi

where X [m, M ]. for i = 1, . . . , n.

unknown
j =1 i=1

aij

ai(j )
j =1 i=1

and
j =1 i=1

aij

ai(j )
j =1 i=1

Samuelson

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

Where =

)2 /n.

Markov

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

where X is a random variable (r.v.), a > 0. for f 0, and f (x) s > 0 for all x S . where t > 0. where t > 0. for X [0, 1] and c 0, E[X ] .

Kolmogorov

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

Pr max |Sk |
k

1 1 Var[Sn ] = 2 2
k i=1

Var[Xi ]
i

where X1 , . . . , Xn are i.r.v., E[Xi ] = 0, Var[Xi ] < for all i, Sk = Etemadi

Chebyshev

Pr X E[X ] t Var[X ]/t2 Pr X > 0 (E[X ])2 /(E[X 2 ]) Var[X ]/(E[X 2 ]) E ( X )k tk

Xi and > 0.

Pr X E[X ] t Var[X ]/(Var[X ] + t2 ) where E[X 2 ]

Pr max |Sk | 3 3 max Pr |Sk | where Xi are i.r.v., Sk =

2nd

moment

Pr X = 0 kth moment

where E[X ] 0.

0. for martingale (Xk ) and > 0.

= 0.

Doob

Pr max1kn |Xk | E |Xn | /

Pr X t

and Bennett Pr

Pr X t Ck X= Cherno Pr[X t] F (a)/at F (z ) =

for Xi [0, 1] k-wise indep. r.v., Xi , i = 1, . . . , n, = E[X ], Ck = 2 ke1/6k 1.0004, k even. for X r.v., Pr[X = k] = pk , e (1 + )(1+) e

i=1

Xi
1 n

exp

n 2 M M2 n 2

where Xi i.r.v.,

E[Xi ] = 0, 2 =

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

Var[Xi ], |Xi | M (w. prob. 1), 0, 2 + M /3)

pk z k probability gen. func., and a 1. exp

Bernstein

Pr
i=1

Pr X (1 + )

2 min{2 + , 3} Xi , = E[X ], 0. 2 2 for [0, 1). Azuma

Xi

exp

2(n 2

for Xi i.r.v.,
1 n

E[Xi ] = 0, |Xi | < M (w. prob. 1) for all i, 2 = Pr Xn X0 2 exp 2 2

Var[Xi ], 0.

where Xi indep. r.v. drawn from {0,1}, X = Pr X (1 ) (1 )(1) exp 2 R

for martingale (Xk ) s.t.

Simpler (weaker) form: Pr X R Pr X t

for R 2e ( 5.44). Efron-Stein Xi .

Xi Xi1 < ci (w. prob. 1), for i = 1, . . . , n, 0. Var[Z ] 1 E 2

pk

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

Z Z (i )

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

f : X n R, Z = f (X1 , . . . , Xn ), Z (i) = f (X1 , . . . , Xi , . . . , Xn ). McDiarmid Pr Z E[Z ] 2 exp 2 2 n 2 i=1 ci for Xi , Xi X i.r.v.,

= /(1 p), E[Xi ] = pi , X = kk Hoeding Pr X E[X ] 2 exp Xi [ai , bi ] (w. prob. 1), X = E eX exp VysochanskijPetunin-Gauss 2 (b 8 a )2
n i=1

Pr X (1 + )

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

> 0.

Xi , 0.

2 2 (b i a i )2

for Xi i.r.v., Janson

Z , Z (i) as before, s.t. Z Z (i) ci for all i, and 0. M Pr M = B i M exp 2 2

where Pr[Bi ] for all i, Pr[Bi Bj ]. (1 xj ),

A related lemma, assuming E[X ] = 0, X [a, b] (w. prob. 1) and R: Lov asz
4 92

(1 Pr[Bi ]), = Bi

Pr

Pr X E[X ] Pr X m
4 2 9 2

if
2 , 3

8 , 3

(1 xi ) > 0

where Pr[Bi ] xi

(i,j )D

if
3

If each Bi mutually indep. of the set of all other events, exc. at most d,
2 . 3

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

Pr X m 1

if

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

B i > 0.

Where X is a unimodal r.v. with mode m, 2 = Var[X ] < , 2 = Var[X ] + (E[X ] m)2 = E[(X m)2 ]. Paley-Zygmund Pr X E[X ] 1 Var[X ] (1 )2 (E[X ])2 + Var[X ] for X 0,

1 1 n2 ln n xk xj 8

where 1 x1 xn 1.

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

L aszl o Kozma latest version: http://www.Lkozma.net/inequalities_cheat_sheet