Useful Inequalities

Cauchy-Schwarz

Minkowski

H
older




n
P

n
P

i=1

n
P

i=1

Bernoulli

xi y i

i=1

2

|xi + yi |p

|xi yi |

n
P

i=1

1

n
P

i=1

x2i



n
P

|xi |p

(1 + x)r 1 + rx

{x2 > 0}

i=1

i=1

1/p 

(1 +

x)r

(1 +

nx)n+1

1+

1

n
P

i=1

|yi |q

n
P

i=1

|yi |p

1/q

1

for p 1.

for p, q > 1,

1
p

1
q

(1 + (n +

for x [1,

1
),
r1

for x R, n N.

ex

xe

1
2x

1+
xn
n!

for x R, and

ex 1 + x +
ex

1 + x,

x2
2

x
2

1)

x2
n

ex 1

+1

ex

1+

for x 0, reverse for x 0.

for x [0, 1.59] and

< xx < x2 x + 1

x1/r (x

x n
n

rx(x1/r

2x

for n > 1, |x| n.

x n+x/2
n

x
2

for x [0, 1].

1)

for x, y > 0.

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

logarithm

x1
x
2x
2+x

ln(x)

x2 1
2x

1
n

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

ln(1 + x)

x
2

ln(1 + x) x
x3
2

hyperbolic

x cos x
2
x

for x, y R.

1
i=1 i

x3
2

Stirling

means

min{xi }

Lehmer

n n
e

2n .

Heinz

Newton

2n

1
x

<

x+1

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

n
1
i xi

xi

x1 < 2 x2 x1

n!

1/n

1
n

2n

n n 1/12n
e
e

xi

1
n

for x 1.

en

n n
e

xi 2 max{xi }

P
P
p 1/p , w 0,
Mp Mq for p q, where Mp =
i
i wi |xi |
i wi = 1.
Q
In the limit M0 = i |xi |wi , M = mini {xi }, M = maxi {xi }.

log mean

wi |xi |p

wi |xi |p1

xy
xy

x+
2

P
wi |xi |q
P i
q1
i wi |xi |

(xy) 4

x1 y +x y 1
2

Sk 2 Sk1 Sk+1
P
1
Sk =  n
k

Jensen

1i1 <<ik n

xy
ln(x)ln(y)
x+y
2

p
k

and

for p q, wi 0.

x+
2

2

x+y
2

for x, y > 0.

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

Sk

(k+1)

ai 1 ai 2 aik ,

Sk+1

and

for 1 k < n,

ai 0.

ln(n) + 1

Chebyshev

x3
3

tan x

sin x,

x2
,
sinh x


all for x 0,

for x R, [1, 1].

P
pi xi i pi (xi )

n
P

f (ai )g(bi )pi

n
P

ai b i

i=1

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

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

where pi 0,

pi = 1, and convex.

n
P

f (ai )pi

i=1

n

P

i=1

n


P
g(bi )pi
f (ai )g(bni+1 )pi
i=1

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




 

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

for x [0, 0.43], reverse elsewhere.

x 3 cos x x x3 /6 x cos

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

for x [0, 0.45], reverse elsewhere.

ex(+x/2)

nn
kk (nk)nk

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

Pn

x cos x
1x2 /3

cosh(x) + sinh(x)

2 x+12 x <

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

x3
4

x2
2

x cos x
x3
sinh2 x

square root

Maclaurin-

for x [0, 2.51], reverse elsewhere.

x2
2

ln(1 x) x
trigonometric

x 1,

x
x+1

ln(1 + x)

for x, n > 0.

for x (0, 1).

for x, r 1.
xy

y
xy + y x > 1 and ex > 1 + x
> e x+y
y

Pd
n

n
for n d 0.
nd + 1 and
i=0 i 2


Pd
d
n

en
for n d 1.
i=0 i
d


Pd
n

n

d
for n
d 0.
i=0 i d 1 + n2d+1
2






P
n
n
n n
1
i=0 i 12 n for (0, 21 ).
n

power means

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


x n
n

n

k

and

n

i=0 i

r > 1.

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

ex 1 +

(en)k
kk

Pd

= 1.

(a + b)n an + nb(a + b)n1 for a, b 0, n N.

p

q
1 + xq
for (i) x > 0, p > q > 0,
1+ x
p

exponential

nk
k!

2n(1)

for x [1, 0], n N.

1)x)n

n

k

n
n

1
1
for n k 0 and 4n (1 8n
4n (1 9n
n
) 2n
).
k
n






n1 +n2
n1 n2
for n1 k1 0, n2 k2 0.
k +k
k1 k2
1
2

nH()

n

, H(x) = log2 (xx (1x)1x ).
G n G for G = 2
2

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

rx
1(r1)x

(nk+1)k
}
k!

max { n
,
kk
nk
4k!

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

1
1nx

(1 +

yi2

|xi |p

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

x)n

n
P

binomial

v0.27a November 29, 2014

rearrangement

i=1

n
P

i=1

ai b(i)

n
P

ai bni+1

i=1

for a1 an ,

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

n
n
n
P
P
P
fi (bi )
fi (b(i) )
fi (bni+1 )
i=1

i=1

i=1

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

pi = 1.

Weierstrass

1 xi

wi

w i xi

where xi 1, and

Carleman

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

P
If wi [0, 1],
wi 1 and xi 1, the reverse holds.
Young

Kantorovich

( px1p

xi 2

1 1
)
qxq

xi
yi

0<m
sum-integral

RU

Cauchy

(a)

L1

Hermite

Chong

n
P

Gibbs

P
Shapiro

i
i

ai

ai log
ai

ai
bi

(b)

Rb
a

ai

n
Q

i=1
a
b
b
a

1
q

= 1.

for f nondecreasing.

(a)+(b)
2
n
Q

ai a i

ai a(i)

Karamata

Muirhead
for ai > 0.

Hadamard

(det A)2

Schur

Pn

i=1

A2ij

i=1 j=1

2i

Pn

Hilbert

i,j=1

A2ij

and

i=1

di

Pk

i=1

Acz
el

Qn
Pn
a
P
x i
a i xi
Qn i=1 i a Pn i=1
for xi [0, 12 ], ai [0, 1],
ai = 1.
i
i=1 (1 xi )
i=1 ai (1 xi )
a1 b 1

given that
Mahler

n
Q

Pn
a21

xi + yi

i=1

Abel
Milne

b1 min
k

Pn

ai b i
Pn

i=2

k
P

i=1

i=1 (ai

>

+ bi )

a21

2
i=2 ai

1/n

ai

2

n
P

i=1

n
Q

i=1

or

1/n
xi

b21

Pn
>

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

n
Q

i=1

ai bi b1 max

P

ai bi
n
i=1 ai +bi

1/n
yi

Pn

2
i=2 bi

i=1

(ai )

1
n!

m=1

Pn

i=1

Pn

i=1 bi

m P
n
Q

j=1 i=1

ai(j)

for a1 a2 an and b1 bn ,
Pt
Pt
i=1 ai
i=1 bi for all 1 t n,

1
n!

n
1
xb(n)
xb(1)

1
c2 +1/2

<

Copson

 P
P

n=1

n=1

Kraft
LYM

kn

4

m=1

n=1

a2n

2n
n=1 (n2 +c2 )2

ak
k

p

2c(i) 1

XA

Sauer-Shelah

a2m

1 P
2

n
1
|X|

pp

n=1

1
c2

<

an p

n=1

n2 a2n

1

2 2
n=1 bn

for am , bn R.

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

for c(i) depth of leaf i of binary tree, sum over all leaves.
1, A 2[n] , no set in A is subset of another set in A.

|A| |str(A)|

Pr
Pr

vc(A)
P
i=0

n

i

for A 2[n] , and

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

n
 W

k

P
Ai
(1)j1 Sj

for 1 k n, k odd,

n
 W

k

P
Ai
(1)j1 Sj

for 2 k n, k even.

i=1

Sk =

for an R.

for c 6= 0.

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

for b1 bn 0.
ai

am bn
n=1 m+n

Mathieu

Bonferroni

ai

j=1 i=1

aij

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

P  a1 +a2 ++an p  p p P
p
p1
for an 0, p > 1.
n=1 an
n=1
n

where xi , yi > 0.

i=1

m P
n
Q

and

(bi )

i=1

i=1

k
P

Pn

n
1

xa
xa
(n)
(1)

an

1 n the eigenvalues, d1 dn the diagonal elements.

Ky Fan

Pn

for 1 k n.

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

ai(j)

j=1 i=1

|ak |

k=1

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

Q n
 Pn
Qn


for |ai |, |bi | 1.
i=1 ai
i=1 bi
i=1 |ai bi |
Qn
Q
n
n
i=1 ( + ai ) (1 + ) , where
i=1 ai 1, ai > 0, > 0.
P
 P
 P
 P

1+y 1y
1y 1+y
1+x 1x
1x 1+x
bi
bi

bi
bi
i ai
i ai
i ai
i ai

Carlson

where A is an n n matrix.
Pk

Pn

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

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

where x, y, z 0, t > 0

n
m Q
P

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

i=1

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

1/k

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

Hardy

n P
n
Q

aij

|ai |

and {ai }  {bi } (majorization), i.e.

and n 12 if even, n 23 if odd.

Schur

m Q
n
P

k
i=1

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

for convex.

for ai , bi 0, or more generally:

P
P
for concave, and a :=
ai , b :=
bi .

n
2

Callebaut
f (x) dx

k=1

j=1 i=1

where a < b, and convex.

(x) dx

a log

bi

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

R U +1

f (i)

and

xi
xi+1 +xi+2

i=1

i=L

a(i)

n
P

PU

1
ba

1
p

for xi , yi > 0,

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

M < ,

f (b)f (a)
ba

yq
q

2

 A 2 P
yi 2 G
i xi yi

f (x) dx

a+b
2

i=1

P

xy

xp
p

sum & product

Q

Pn

j=1

j=1

1i1 <<ik n

Pr Ai1 Aik

where Ai are events.

Bhatia-Davis
Samuelson

Markov

Chebyshev

2nd moment

kth moment

Chernoff

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

where X [m, M ].

n 1 xi + n 1 for i = 1, . . . , n.
P
P
Where =
xi /n , 2 =
(xi )2 /n.




E (X )
and
Pr X  t
tk

k/2



nk
for Xi [0, 1] k-wise indep. r.v.,
Pr X  t Ck
t2

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

Pr[X t] F (a)/at for X r.v., Pr[X = k] = pk ,

P
F (z) = k pk z k probability gen. func., and a 1.






e
2
Pr X (1 + )

exp
min{2 + , 3}
(1 + )(1+)
P
where Xi indep. r.v. drawn from [0,1], X =
Xi , = E[X], 0.






e
2
Pr X (1 )
for [0, 1).

exp
2
(1 )(1)


Simpler (weaker) form: Pr X R 2R for R 2e ( 5.44).


Pr X t


 
n
k

pk
 
t
k

Pr X (1 + )

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

(1+)

n
k
p /

k
k

2 2
Pn
2
i=1 (bi ai )
P
Xi [ai , bi ] (w. prob. 1), X =
Xi , 0.



Pr X E[X] 2 exp

,
n

q



8
if
Pr X E[X] 94 2
,
3



2
2
4
if ,
Pr X m 92
3




2


Pr X m 1
if
.
3

Etemadi

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






Pr max |Sk | 3 3 max Pr |Sk |
1kn
1kn
P
where Xi are i.r.v., Sk = ki=1 Xi , 0.




Pr max1kn |Xk | E |Xn | /

Doob

Bennett

Bernstein

Azuma

Efron-Stein

McDiarmid

> 0.
Janson

for Xi i.r.v.,
Lov
asz

for martingale (Xk ) and > 0.



n 2
M

exp 2
where Xi i.r.v.,
2
M
n
i=1
P
1
Var[Xi ], |Xi | M (w. prob. 1), 0,
E[Xi ] = 0, 2 = n
Pr

n
 P

Xi

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


n
 P

2
Pr
Xi exp
2
2(n + M /3)
i=1

for Xi i.r.v.,

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





2
P
Pr Xn X0  2 exp
for martingale (Xk ) s.t.
2
2 n
i=1 ci


Xi Xi1  < ci (w. prob. 1), for i = 1, . . . , n, 0.

Var[Z]

1
2

n
P

i=1

Z Z (i)

2

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

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






2 2
Pr Z E[Z] 2 exp Pn
for Xi , Xi X i.r.v.,
2
i=1 ci


Z, Z (i) as before, s.t. Z Z (i)  ci for all i, and 0.

V 
B i M exp
M Pr
Q
M = (1 Pr[Bi ]), =

2 2
P

i6=j,Bi Bj

Pr

V

 Q
B i (1 xi ) > 0

where Pr[Bi ] for all i,

Pr[Bi Bj ].

where Pr[Bi ] xi

(i,j)D

(1 xj ),

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

If each Bi mutually indep. of the set of all other events, exc. at most d,
V 
B i > 0.
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,

P
Var[Xi ] < for all i, Sk = ki=1 Xi and > 0.

for X 0,

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

Xi .

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


 2


(b a)2
E eX exp
8


P
1
Pr max |Sk | 2 Var[Sn ] = 12
Var[Xi ]
k

Petunin-Gauss

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

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

kk

Kolmogorov

Vysochanskij-


k

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



Pr X E[X] 1

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





Pr |X| a E |X| /a where X is a random variable (r.v.), a > 0.




Pr X c (1 E[X])/(1 c) for X [0, 1] and c 0, E[X] .

Pr X S] E[f (X)]/s for f 0, and f (x) s > 0 for all x S.



Pr X E[X] t Var[X]/t2 where t > 0.


Pr X E[X] t Var[X]/(Var[X] + t2 ) where t > 0.


Pr X > 0 (E[X])2 /(E[X 2 ]) where E[X] 0.


Pr X = 0 Var[X]/(E[X 2 ]) where E[X 2 ] 6= 0.

Hoeffding

Paley-Zygmund

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