` n n ` n n 1/(12n+1) ` n n 1/12n ` n n

Useful Inequalities {x2 > 0} Stirling e

e
2n
e
e n! 2n
e
e en
e
version 0.10a August 1, 2011
x3 x cos x x3 x
n
!2 n
! n
! trigonometric x x cos x 2
x 3 cos x x x cos sin x
X X X 2 1 x /3 6 3
Cauchy-Schwarz x i yi x2i yi2
x3 2
x sin x, for x 0, 2 .

i=1 i=1 i=1 x cos(0.56 x) x x + tan x, and
3
n
!1 n
!1 n
!1 2
p p p
X
p
X
p
X
p hyperbolic cosh(x) + sinh(x) ex /2+x , where x R, [1, 1].
Minkowski |xi + yi | |xi | + |yi | , for p 1.
i=1 i=1 i=1
a+b ba
!1/p !1/q Napier b> > > ab > a, for 0 < a < b.
n
X n
X n
X 1 1 2 ln(b) ln(a)
H
older |xi yi | |xi |p |yi |q , for p, q > 1, + = 1.
i=1 i=1 i=1
p q sP
xi 2
P
xi Y 1/n n
means max{xi } xi P 1 min{xi }
r n n xi
Bernoulli (1 + x) 1 + rx, for x > 1, r R \ (0, 1). If r = 2n (n N),
inequality holds for x R. Reverse holds for x > 1, r [0, 1]. power means r Ms ,
Mw for all pairs r s, where:
w
rx 1 !1/r
(1 + x)r 1 + , for r > 1, 1 x < . n
1 (r 1)x r1
X
r (x , x , . . . , x ) = r P
Mw 1 2 n w x
i i and wi = 1.
x n i=1
ex 1 +
`
exponential 1 + x, for x R, n > 0 (left), n 1 (right). r tends to min, geom. mean and max, respectively.
n If r = , 0, +, Mw
If x < 0, then n x is required for both. Outer inequality always holds. pk
p
Maclaurin Sk (k+1) Sk+1 , for 1 k < n, where:
x2
ex xe , for x R, ex 1 + x + , for x 0, reverse for x 0. 1 X
2 Sk = `n ai1 ai2 aik , and ai > 0.
k 1i1 <<ik n
x x2 x3
logarithm log(1 + x) min x, x + , for x > 1.
x+1 2 3 Newton 2
Sk Sk1 Sk+1 , for 1 k < n, and Sk as before.
2x x
log(1 + x) , for x 0. Reverse for x (1, 0].
2+x x+1 Jensen (E [X]) E [(X)], where X is a random variable, and convex.
x2 x3 For concave the reverse holds. Without probabilities:
log(1 + x) x + , for x [0, 0.45], reverse elsewhere.
2 4 n
`X Xn
P
x2 x3 pi xi pi (xi ), where pi 0, pi = 1.
log(1 x) x , for x [0, 0.43], reverse elsewhere. i=1 i=1
2 2
n n n n n
1
X `X `X X
Chebyshev f (ai )g(bi )pi f (ai )pi g(bi )pi f (ai )g(bni+1 )pi ,
X
harmonic log(n + 1) log(n) + 1
i=1
i i=1 i=1 i=1 i=1
P
for a1 an , b1 bn , and f, g nondecreasing, pi 0, pi = 1.
1
square root 2 x + 1 2 x < < 2 x 2 x 1, for x 1.
x With expectations: E f (X)g(X) E f (X) E g(X) .
n n n
` n k n nk ` en k X X X
binomial , for n k > 0. rearrangement ai bi ai b(i) ai bni+1 , for a1 an ,
k k k! k i=1 i=1 i=1
d b1 bn and a permutation of [n]. More generally:
X n
binomial sum nd + 1, for n d 0, n n n
i X X X
i=0 fi (bi ) fi (b(i) ) fi (bni+1 ),
d i=1 i=1 i=1
X n ` en d
, for n d 1.
`
with fi+1 (x) fi (x) nondecreasing for all 1 i < n.
i=0
i d
xp yq 1 1
22n 2n 22n Young xy + , for x, y 0 and p, q > 0 , + = 1.
middle binomial p q p q
2 n n 2n
n n n
ai
ai a(i) ,
X Y Y
binomial ratio
` n
(1 )(1)
n
, for (0, 1). Chong n, and ai ai for ai > 0.
n a
i=1 (i) i=1 i=1
 n
! n
! 2 n
!2 n n n
X X A X Y ` 1/n Y 1/n
Y 1/n
Kantorovich xi 2 yi 2 xi y i , for xi , yi > 0, Mahler xi + yi xi + yi , where xi , yi > 0.
i=1 i=1
G i=1 i=1 i=1 i=1
xi
0<m M < , A = (m + M )/2, G= mM . n
m Y n
m Y n
m X n
m X
yi X X Y Y
unknown aij ai(j) , and aij ai(j) ,
f (b) f (a) j=1 i=1 j=1 i=1 j=1 i=1 j=1 i=1
Cauchy (a)
(b), where a < b, and convex.
ba for 0 ai1 aim for i = 1, . . . , n and is a permutation of [n].
For concave the reverse holds.
n
X n
X

a+b

1
Z b (a) + (b) Karamata (ai ) (bi ), where a1 a2 an and b1 bn ,
Hadamard (x) dx , for convex.
2 ba a 2 i=1 i=1
t t
X X
n and {ai }  {bi } (majorization), i.e. ai bi for all 1 t n,
X ai a P P i=1 i=1
Gibbs ai log a log , for ai , bi 0, a := ai , b := bi .
i=1
bi b with equality for t = n and is convex (for concave the reverse holds).

n
` bi `b 1 X a1 1 X b1
xa xb(n)
X
n n
Woeginger ai a , for concave and variables as before. Muirhead x x ,
i=1
ai a n! (1) (n) n! (1)

` x p ` x q where a1 a2 an and b1 b2 bn and {ak }  {bk },

Pe
cari
c 1+ 1+ , where either (i) x > 0, p > q > 0,
p q xi 0 and the sums extend over all permutations of [n].
(ii) p < q < x < 0 or (iii) q > p > x > 0. Reverse, if
n k
!1/k n
(iv) q < 0 < p , q > x > 0 or (v) q < 0 < p , p < x < 0. X Y X
Carleman |ai | e |ak |
n k=1 i=1 k=1
X xi n
Shapiro , where xi > 0, (xn+1 , xn+2 ) := (x1 , x2 ),
i=1
xi+1 + xi+2 2
n n n n
! ! ! !
X X ai bi X X
and n 12 if even, n 23 if odd. Milne (ai + bi ) ai bi
i=1
a + bi
i=1 i i=1 i=1
Schur xt (x y)(x z) + y t (y z)(y x) + z t (z x)(z y) 0,
where x, y, z 0, t > 0 k
X n
X k
X
Abel bn min |ai | |ai bi | bn max |ai |, for 0 b1 bn .
n n k k
Y ` w X i=1 i=1 i=1
Weierstrass 1 xi i 1 w i xi , where xi 1, and
i=1 i=1 !1 !1
X 2 2
either wi 1 (for all i) or wi 0 (for all i).
X am bn X X
Hilbert a2m b2n , for am , bn R.
P m+n
If wi [0, 1], wi 1, and xi 1, the reverse holds. m=1 n=1 m=1 n=1
If we put max{m, n} instead of m + n, we have 4 instead of .
n n
a
Y X
xi i a i xi

a1 + a2 + + an p
p X
i=1 i=1 p
for xi [0, 21 ], ai [0, 1],
X
apn ,
P
Ky Fan n n , ai = 1. Hardy for an 0, p > 1.
Y
ai
X
n=1
n p1 n=1
(1 xi ) ai (1 xi )
i=1 i=1

!4
n n n
X X X
X 2 X X Carlson an 2 an 2 n2 an 2 , for an R.
a21 a2i b21 b2i ,
` ` `
Acz
el a1 b1 ai bi n=1 n=1 n=1
i=2 i=2 i=2
n
X Xn

given that a21 > a2i or b21 > b2i . 1 X 2n 1
i=2 i=2
Mathieu < < 2, for c 6= 0.
c2 + 1/2 n=1
(n2 + c2 )2 c
n n n n
! ! ! !
a1+y b1y a1y b1+y
X
a1+x b1x
X
a1x b1+x
X X
Callebaut i i i i i i i i ,
0 1p

X X ak X
i=1 i=1 i=1 i=1 A pp
Copson @ an p , for an 0, p > 1, reverse if p (0, 1).
for 1 x y 0. n=1 kn
k n=1

n k E |Xn |
" #

Pr max |Xk | , for martingale (Xk ) and > 0.
[ X
Bonferroni Pr Ai (1)j1 Sj , for 1 k n, k odd, Doob
1kn
i=1 j=1
n
2

n k
" # X
Pr
[
Ai
X
(1) j1
Sj , for 2 k n, k even. Bernstein Pr Xi exp , for Xi independent,
i=1
2(n 2 + M /3)
i=1 j=1
1
E[Xi ] = 0, |Xi | < M (w. prob. 1) for all i, 2 =
P
X n
Var[Xi ], 0.
Sk = Pr Ai1 Aik , where Ai are events.

1i1 <<ik n e 2


Chernoff Pr X (1 + ) exp ,
(1 + )(1+) min{2 + , 3}
E |X|
Pr |X| a , where X is a random variable, a > 0.
P
Markov where Xi independently drawn from {0,1}, X = Xi , = E[X], 0.
a
e 2

1 E[X]
Pr X (1 ) exp , for [0, 1).

Pr X c , for X [0, 1] and c 0, E[X] .
1c (1 )(1) 2
n
Without probabilities: , where c is the number of
c
Simpler (weaker) form: Pr X R 2R , for R 2e ( 5.44).

a
elements a, among n nonnegative numbers with mean .
2 2


Hoeffding Pr X E[X] 2 exp Pn 2
, for Xi indep.
Var[X] i=1 (bi ai )
Chebyshev Pr X E[X] t , P
t2 Xi [ai , bi ] (w. prob. 1), X = Xi , 0.
Var[X] A related lemma, assuming E[X] = 0, X [a, b] (w. prob. 1) and R:
Pr X E[X] t , where t > 0 (for both).
Var[X] + t2 2
(b a)2

E eX exp

` ` 8
Bhatia-Davis Var[X] M E[X] E[X] m , where X [m, M ].
2


Azuma Pr Xn X0 2 exp Pn 2
, for martingale (Xk ) s.t.
Samuelson n 1 xi + n 1, for i = 1, . . . , n. 2 i=1 ci

Xi Xi1 < ci (w. probability 1), for i = 1, . . . , n, 0.
xi /n , 2 = (xi )2 /n.
P P
Where =
" n #
r 1 X` 2
4 8 Efron-Stein Var[Z] E Z Z (i) , for Xi , Xi X independent,
Vysochanskij- Pr X E[X] , if , 2
Petunin-Gauss 92 3 i=1

4 2 2 f : X n R, Z = f (X1 , . . . , Xn ), Z (i) = f (X1 , . . . , Xi , . . . , Xn ).

Pr X m , if ,
92 3

2 2

, for Xi , Xi X indep.,

2 McDiarmid Pr Z E[Z] 2 exp Pn 2
Pr X m 1 , if . i=1 ci
3 3
Z, Z (i) as before, s.t. Z Z (i) ci for all i, and 0.

Where X is a unimodal random variable with mode m,

2 = Var[X] < , 2 = Var[X] + (E[X] m)2 = E[(X m)2 ].
Janson M Pr no Bi occurs M exp , where Pr[Bi ] for all i,
2 2
1 1 X
Kolmogorov Pr max |Sk | 2 Var[Sn ] = 2 Var[Xi ], Y X
k i M = (1 Pr[Bi ]), = Pr[Bi Bj ], for Bi and Bj dependent.
i6=j
where X1 , . . . , Xn are independent random variables, E[Xi ] = 0,
1 1
Var[Xi ] < for all i, Sk = ki=1 Xi and > 0.
X
n2 log n,
P
Erd
os where 1 x1 xn 1.
1j<kn
xk xj 8
`
Etemadi Pr max |Sk | 3 3 max Pr |Sk | ,
1kn 1kn N
X
Pk Kraft 2c(i) 1, where N is the number of leaves in a binary tree,
where Xi are independent random variables, Sk = i=1 Xi , 0. i=1
n and c(i) is the depth of a leaf i.
n 2 ` M

X
Bennett Pr Xi exp 2 , where Xi independent,
i=1
M n 2
1 P
E[Xi ] = 0, 2 = n Var[Xi ], |Xi | M (w. probab. 1), 0, c b a L
aszl
o Kozma (www.Lkozma.net)
References and latest version of this file:
(u) = (1 + u) log(1 + u) u. http://www.Lkozma.net/inequalities_cheat_sheet