√ √ √ √ √ √

√1
Useful Inequalities {x2 > 0} v0.29 · June 17, 2017 square root 2 x+1−2 x < x
< x+1− x−1 < 2 x−2 x−1 for x ≥ 1.

x x2 √ x
 n
2  n
 n
 1− 2
− 2
≤ 1−x≤1− 2
.
x2i yi2
P P P
Cauchy-Schwarz xi y i ≤
i=1 i=1 i=1 k (n−k+1)k n
nk (en)k n
nn
binomial max { n
kk
, k!
} ≤ k
≤ k!
≤ kk
and k
≤ kk (n−k)n−k
≤ 2n .
1 1 1
n n n √
  
p p p
nk n n
|xi + yi |p |xi |p |yi |p ≤ n for n ≥ k ≥ 0 and √4πn (1 − 8n1
) ≤ 2n 1
≤ √4πn (1 − 9n



).
P P P
Minkowski ≤ + for p ≥ 1. 4k! k n
i=1 i=1 i=1
n1 n2
≤ nk1 +k
+n2




1/p  1/q k1 k2
for n1 ≥ k1 ≥ 0, n2 ≥ k2 ≥ 0.
n
 n n 1 2
1 1 √
|xi |p |yi |q
P P P
Hölder |xi yi | ≤ for p, q > 1, + = 1. π n nH(α)
≤ G for G = √ 2 , H(x) = − log2 (xx (1−x)1−x ).


i=1 i=1 i=1
p q
2
G ≤ αn
2πnα(1−α)

Bernoulli (1 + x)r ≥ 1 + rx for x ≥ −1, r ∈ R \ (0, 1). Reverse for r ∈ [0, 1]. Pd n

≤ nd +1 and
Pd n

≤ 2n for n ≥ d ≥ 0.
i=0 i i=0 i
(1 + x)r ≤1+ (2r − 1)x for x ∈ [0, 1], r ∈ R \ (0, 1). Pd n
en d


i=0i d
≤ for n ≥ d ≥ 1.
1
(1 + x)n ≤ 1−nx
for x ∈ [−1, 0], n ∈ N. Pd n
n
d

n
i=0 i ≤ d 1 + n−2d+1 for 2
≥ d ≥ 0.
rx 1
(1 + x)r ≤ 1 + for x ∈ [−1, ), r > 1. n
n
n

1−(r−1)x r−1 ≤ αn 1−α
for α ∈ (0, 21 ).
P
αn i=0 i ≤ 1−2α αn
(1 + nx)n+1 ≥ (1 + (n + 1)x)n for x ∈ R, n ∈ N. √ √
n n n n 1/(12n+1) n n 1/12n n n





xy > x
for x > 0, y ∈ (0, 1). Stirling e e
≤ 2πn e
e ≤ n! ≤ 2πn e
e ≤ en e
x+y

(a + b)n ≤ an + nb(a + b)n−1 for a, b ≥ 0, n ∈ N. q P 2

x
n
xi )1/n ≤ 1 1
xi 2 ≤
Q P P
means min xi ≤ ≤( xi ≤ P i ≤ max xi
P −1
xi n n xi
x n x n x2
ex ≥ 1 + ≥ ex 1 −




exponential n
≥ 1 + x, 1+ n n
for n > 1, |x| ≤ n.
xn x n+x/2 p 1/p , w ≥ 0,
P
P
e x ≥ xe + 1 ≤ ex ≤ 1 + power means Mp ≤ Mq for p ≤ q, where Mp = i wi |xi | i wi = 1.


for x ∈ R, and n! n
for x, n > 0. i
In the limit M0 = i |xi |wi , M−∞ = mini {xi }, M∞ = maxi {xi }.
Q
x2
ex ≥ 1 + x + 2
for x ≥ 0, reverse for x ≤ 0.
wi |xi |p q
P P
i wi |xi |
x x
e−x ≤ 1 − for x ∈ [0, ∼ 1.59] and 2−x ≤ 1 − for x ∈ [0, 1].
2 2 Lehmer P i p−1
≤ P q−1
for p ≤ q, wi ≥ 0.
1
< xx < x2 −x+1 for x ∈ (0, 1). i wi |xi | i wi |xi |
2−x
√ √ √  1
√ √ 2
x1/r (x − 1) ≤ rx(x1/r − 1)
for x, r ≥ 1. log mean xy ≤
x+ y
(xy) 4 ≤ x−y
≤
x+ y
≤ x+y
for x, y > 0.
2 ln(x)−ln(y) 2 2

y xy
x + y > 1 and e > 1 + x
y x x
y
> e x+y for x, y > 0.
√ x1−α y α +xα y 1−α x+y
2 Heinz xy ≤ 2
≤ 2
for x, y > 0, α ∈ [0, 1].
2−y− e−x−y
≤1+x≤y+ and ≤ x + ex ex−y , ex for x, y ∈ R.
p q
1+ x ≥ 1 + xq



for (i) x > 0, p > q > 0,
p
Sk 2 ≥ Sk−1 Sk+1 k (k+1)
p
p Maclaurin- and Sk ≥ Sk+1 for 1 ≤ k < n,
(ii) − p < −q < x < 0, (iii) − q > −p > x > 0. Reverse for: 1 P
Newton Sk =  n ai 1 ai 2 · · · aik , and ai ≥ 0.
k 1≤i1 <···<ik ≤n
(iv) q < 0 < p , −q > x > 0, (v) q < 0 < p , −p < x < 0.
P
P P
Jensen ϕ i pi xi ≤ i pi ϕ (xi ) where pi ≥ 0, pi = 1, and ϕ convex.
x−1 x2 −1 1
logarithm x
≤ ln(x) ≤ 2x
≤ x − 1, ln(x) ≤ n(x n − 1) for x, n > 0.
Alternatively: ϕ (E [X]) ≤ E [ϕ(X)]. For concave ϕ the reverse holds.
2x √x
2+x
≤ ln(1 + x) ≤ x+1
for x ≥ 0, reverse for x ∈ (−1, 0].
n
P n
P n

P
n
P
1 Pn 1 Chebyshev f (ai )g(bi )pi ≥ f (ai )pi g(bi )pi ≥ f (ai )g(bn−i+1 )pi
ln(n + 1) < ln(n) + n
≤ i=1 i ≤ ln(n) + 1 i=1 i=1 i=1 i=1
x P
ln(1 + x) ≥ 2
for x ∈ [0, ∼ 2.51], reverse elsewhere. for a1 ≤ · · · ≤ an , b1 ≤ · · · ≤ bn and f, g nondecreasing, pi ≥ 0, pi = 1.
     
x2 x3 Alternatively: E f (X)g(X) ≥ E f (X) E g(X) .
ln(1 + x) ≥ x − 2
+ 4
for x ∈ [0, ∼ 0.45], reverse elsewhere.
x2 x3 n n n
ln(1 − x) ≥ −x − 2
− 2
for x ∈ [0, ∼ 0.43], reverse elsewhere. rearrangement
P
ai b i ≥
P
ai bπ(i) ≥
P
ai bn−i+1 for a1 ≤ · · · ≤ an ,
i=1 i=1 i=1
x3 x cos x √ x
trigonometric x− 2
≤ x cos x ≤ 1−x2 /3
≤x 3
cos x ≤ x − x3 /6 ≤ x cos √ ≤ sin x, b1 ≤ · · · ≤ bn and π a permutation of [n]. More generally:
3
3 2 n n n
x 2 x
x cos x ≤ sinh 2 x ≤ x cos (x/2) ≤ sin x ≤ (x cos x + 2x)/3 ≤ sinh x ,
P P P
hyperbolic fi (bi ) ≥ fi (bπ(i) ) ≥ fi (bn−i+1 )
i=1 i=1 i=1
2 2 2
n o
max π2 , π −x
≤ sinx x ≤ cos x2 ≤ 1 ≤ 1 + x3 ≤ tan x π
 
for x ∈ 0, .


π 2 +x2 x 2 with fi+1 (x) − fi (x) nondecreasing for all 1 ≤ i < n.
 Q
wi P 1/k
Weierstrass 1 − xi ≥1− w i xi where xi ≤ 1, and Pn Q
i i k Pn
Carleman k=1 i=1 |ai | ≤e k=1 |ak |
either wi ≥ 1 (for all i) or wi ≤ 0 (for all i).
m Q
n m Q
n m P
n m P
n
P
If wi ∈ [0, 1], wi ≤ 1 and xi ≤ 1, the reverse holds. P P Q Q
sum & product aij ≥ aiπ(j) and aij ≤ aiπ(j)
 2 P j=1 i=1 j=1 i=1 j=1 i=1 j=1 i=1
P 2

P 2

A

2
Kantorovich i xi i yi ≤ G i xi yi
for xi , yi > 0, where 0 ≤ ai1 ≤ · · · ≤ aim for i = 1, . . . , n and π is a permutation of [n].
xi √
0<m≤ ≤ M < ∞, A = (m + M )/2, G = mM .
Q n Qn  Pn
yi i=1 ai − i=1 bi ≤ i=1 |ai − bi | for |ai |, |bi | ≤ 1.
 
Qn n
Qn
i=1 (α + ai ) ≥ (1 + α) , where i=1 ai ≥ 1, ai > 0, α > 0.
RU PU R U +1
sum & integral L−1 f (x) dx ≤ i=L f (i) ≤ L f (x) dx for f nondecreasing.
P  P  P  P 
1+x 1−x 1−x 1+x 1+y 1−y 1−y 1+y
f (b)−f (a)
Callebaut i ai bi i ai bi ≥ i ai bi i ai bi
Cauchy ϕ′ (a) ≤ b−a
≤ ϕ′ (b) where a < b, and ϕ convex.
for 1 ≥ x ≥ y ≥ 0, and i = 1, . . . , n.
 
a+b 1
Rb ϕ(a)+ϕ(b) Pn Pn
Hermite ϕ 2
≤ b−a a ϕ(x) dx ≤ 2
for ϕ convex. Karamata i=1 ϕ(ai ) ≥ i=1 ϕ(bi ) for a1 ≥ a2 ≥ · · · ≥ an and b1 ≥ · · · ≥ bn ,
Pt Pt
and {ai }  {bi } (majorization), i.e. i=1 ai ≥ i=1 bi for all 1 ≤ t ≤ n,
P ai a
Gibbs i ai log bi
≥ a log b
for ai , bi ≥ 0, or more generally: with equality for t = n and ϕ is convex (for concave ϕ the reverse holds).
P bi
b

P P
ai ϕ ≤aϕ for ϕ concave, and a = ai , b = bi .
i ai a 1
xa 1
· · · xa 1
xbπ(1)
1
· · · xbπ(n)
P n
P n
Muirhead n! π π(1) π(n)
≥ n! π
n n n
Chong
P ai
≥n and
Q
ai a i ≥
Q
ai aπ(i) for ai > 0. where a1 ≥ a2 ≥ · · · ≥ an and b1 ≥ b2 ≥ · · · ≥ bn and {ak }  {bk },
aπ(i)
i=1 i=1 i=1 xi ≥ 0 and the sums extend over all permutations π of [n].
Schur xt (x − y)k (x − z)k + y t (y − z)k (y − x)k + z t (z − x)k (z − y)k ≥ 0 P∞ P∞ am bn P∞
1 P∞
1
Hilbert m=1 n=1 m+n ≤π m=1 a2m 2 2 2
n=1 bn for am , bn ∈ R.
where x, y, z, t, k ≥ 0.
With max{m, n} instead of m + n, we have 4 instead of π.
xp yq
Young ( px1p + 1 −1
qxq
) ≤ xy ≤ p
+ q
for x, y ≥ 0, p, q > 0, 1
p
+ 1
q
= 1. P∞  a1 +a2 +···+an p  p p P∞ p
Hardy n=1 n
≤ p−1 n=1 an for an ≥ 0, p > 1.
n
P xi n
Shapiro ≥ where xi > 0, (xn+1 , xn+2 ) := (x1 , x2 ), P∞
4 P∞ P∞
i=1
xi+1 +xi+2 2 Carlson n=1 an ≤ π2 n=1 a2n n=1 n2 a2n for an ∈ R.
and n ≤ 12 if even, n ≤ 23 if odd.
1 P∞ 2n 1
Mathieu c2 +1/2
< n=1 (n2 +c2 )2 < c2
for c 6= 0.
n P
n
(det A)2 A2ij
Q
Hadamard ≤ where A is an n × n matrix.
i=1 j=1 n
−1
≤ 1, A ⊂ 2[n] , no set in A is subset of another set in A.
P
LYM |X|
Pn Pn Pk Pk X∈A
Schur i=1 λ2i ≤ i,j=1 A2ij and i=1 di ≤ i=1 λi for 1 ≤ k ≤ n.
A is an n × n matrix. For the second inequality A is symmetric. vc(A)
n

for A ⊆ 2[n] , and
P
Sauer-Shelah |A| ≤ |str(A)| ≤ i
λ1 ≥ · · · ≥ λn the eigenvalues, d1 ≥ · · · ≥ dn the diagonal elements. i=0
Qn a Pn str(A) = {X ⊆ [n] : X shattered by A}, vc(A) = max{|X| : X ∈ str(A)}.
x i a i xi
Qn i=1 i a ≤ Pn i=1 for xi ∈ [0, 12 ], ai ∈ [0, 1],
P
Ky Fan ai = 1.
n k
i=1 (1 − xi ) i=1 ai (1 − xi )
i
(−1)j−1 Sj
 W  P
Bonferroni Pr Ai ≤ for 1 ≤ k ≤ n, k odd,
Pn
2 Pn Pn i=1 j=1
a21 2 b2 2



Aczél a1 b 1 − i=2 ai b i ≥ − i=2 ai 1 − i=2 bi n k
(−1)j−1 Sj
Pn Pn  W  P
given that a21 > i=2 a2i or b21 > 2
i=2 bi .
Pr Ai ≥ for 2 ≤ k ≤ n, k even.
i=1 j=1
n
1/n n n  
1/n 1/n
P
Mahler
Q
xi + yi ≥
Q
xi +
Q
yi where xi , yi > 0. Sk = Pr Ai1 ∧ · · · ∧ Aik where Ai are events.
i=1 i=1 i=1 1≤i1 <···<ik ≤n

k
P n
P k
P Bhatia-Davis Var[X] ≤ M − E[X] E[X] − m where X ∈ [m, M ].
Abel b1 min ai ≤ ai bi ≤ b1 max ai for b1 ≥ · · · ≥ bn ≥ 0.
k i=1 i=1 k i=1
√ √
Pn
 Pn ai bi
 Pn
Pn
Samuelson µ − σ n − 1 ≤ xi ≤ µ + σ n − 1 for i = 1, . . . , n.
Milne i=1 (ai + bi ) i=1 ai +bi ≤ i=1 ai i=1 bi
xi /n , σ 2 = (xi − µ)2 /n.
P P
Where µ =
    
Markov Pr |X| ≥ a ≤ E |X| /a where X is a random variable (r.v.), a > 0.   Var[X]
Paley-Zygmund Pr X ≥ µ E[X] ≥ 1 − for X ≥ 0,
  
Pr X ≤ c ≤ (1 − E[X])/(1 − c) for X ∈ [0, 1] and c ∈ 0, E[X] .
 (1 − µ)2 (E[X])2 + Var[X]

Pr X ∈ S] ≤ E[f (X)]/s for f ≥ 0, and f (x) ≥ s > 0 for all x ∈ S. Var[X] < ∞, and µ ∈ (0, 1).
q
Pr X − E[X] ≥ λσ ≤ 9λ4 2 8
  
Chebyshev Pr X − E[X] ≥ t ≤ Var[X]/t2 where t > 0.
   Vysochanskij- if λ ≥ 3
,
2
4τ 2τ
  
Pr X − E[X] ≥ t ≤ Var[X]/(Var[X] + t2 ) where t > 0.
 
Petunin-Gauss Pr X − m ≥ ε ≤ 9ε2 if ε ≥ √ ,
3
ε 2τ
  
Pr X − m ≥ ε ≤ 1 − √ if ε ≤ √ .
2nd moment Pr X > 0 ≥ (E[X])2 /(E[X 2 ]) where E[X] ≥ 0.
   
3τ 3

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

  Where X is a unimodal r.v. with mode m,
σ 2 = Var[X] < ∞, τ 2 = Var[X] + (E[X] − m)2 = E[(X − m)2 ].
E (X − µ)k
 
kth moment
      

Pr X − µ ≥ t ≤ and Etemadi Pr max |Sk | ≥ 3α ≤ 3 max Pr |Sk | ≥ α
tk 1≤k≤n 1≤k≤n
k/2
where Xi are i.r.v., Sk = ki=1 Xi , α ≥ 0.

nk
P
  
Pr X − µ ≥ t ≤ Ck for Xi ∈ [0, 1] k-wise indep. r.v.,
t2
√
   
Doob Pr max1≤k≤n |Xk | ≥ ε ≤ E |Xn | /ε for martingale (Xk ) and ε > 0.
Xi , i = 1, . . . , n, µ = E[X], Ck = 2 πke1/6k ≤ 1.0004, k even.
P
X=
n nσ 2
 
 
3/2 Mε

E X2
 P 
Bennett Pr Xi ≥ ε
≤ exp − 2 θ where Xi i.r.v.,
2nd 4th where 0 < E X 4 < ∞.
   
and E |X| ≥ M nσ 2

1/2 i=1
4
E [X ] 1
E[Xi ] = 0, σ 2 = n
P
Var[Xi ], |Xi | ≤ M (w. prob. 1), ε ≥ 0,
σ 
Pr X ≥ √ > 0 where E[X] = 0, E[X 2 ] = σ 2 , 0 < E[X 4 ] ≤ tσ 4 .

2 t θ(u) = (1 + u) log(1 + u) − u.
n −ε2
 
Pr[X ≥ t] ≤ F (a)/at for X r.v., Pr[X = k] = pk ,
 P 
Chernoff Bernstein Pr Xi ≥ ε ≤ exp for Xi i.r.v.,
2
2(nσ + M ε/3)
i=1
F (z) = k pk z k probability gen. func., and a ≥ 1.
P
1 P
µ E[Xi ] = 0, |Xi | < M (w. prob. 1) for all i, σ 2 = n Var[Xi ], ε ≥ 0.
eδ −µδ 2
  
 
Pr X ≥ (1 + δ)µ ≤ ≤ exp 
−δ 2

(1 + δ)(1+δ) 3
  
Azuma Pr Xn − X0  ≥ δ ≤ 2 exp for martingale (Xk ) s.t.
2 n 2
P
i=1 ci
P
for Xi i.r.v. from [0,1], X = Xi , µ = E[X], δ ≥ 0 resp. δ ∈ [0, 1).  
µ Xi − Xi−1  < ci (w. prob. 1), for i = 1, . . . , n, δ ≥ 0.
e−δ −µδ 2
  
 
Pr X ≤ (1 − δ)µ ≤ ≤ exp for δ ∈ [0, 1).
(1 − δ)(1−δ) 2 n
 
1

2
Z − Z (i) for Xi , Xi ′ ∈ X i.r.v.,
P
Efron-Stein Var[Z] ≤ 2
E
Further from the mean: Pr X ≥ R ≤ 2−R for R ≥ 2eµ (≈ 5.44µ).
 
i=1
 
n
pk
f : X n → R, Z = f (X1 , . . . , Xn ), Z (i) = f (X1 , . . . , Xi ′ , . . . , Xn ).
k
  P
Pr X ≥ t ≤  
t
for Xi ∈ {0, 1} k-wise i.r.v., E[Xi ] = p, X = Xi .
−2δ 2
 
k
for Xi , Xi ′ ∈ X i.r.v.,
  
McDiarmid Pr Z − E[Z] ≥ δ ≤ 2 exp Pn 2
  n
k̂ (1+δ)µ
i=1 ci
Pr X ≥ (1 + δ)µ ≤ p / for Xi ∈ [0, 1] k-wise i.r.v.,
Z, Z (i) as before, s.t. Z − Z (i)  ≤ ci for all i, and δ ≥ 0.
 
k̂ k̂
P µ
k ≥ k̂ = ⌈µδ/(1 − p)⌉, E[Xi ] = pi , X = Xi , µ = E[X], p = n , δ > 0.  
V  ∆
Janson M ≤ Pr B i ≤ M exp where Pr[Bi ] ≤ ε for all i,
  

−2δ 2
 2 − 2ε
Hoeffding Pr X − E[X] ≥ δ ≤ 2 exp Pn for Xi i.r.v., Q P
2 M = (1 − Pr[Bi ]), ∆ = Pr[Bi ∧ Bj ].
i=1 (bi − ai )
P i6=j,Bi ∼Bj
Xi ∈ [ai , bi ] (w. prob. 1), X = Xi , δ ≥ 0.
V  Q Q
A related lemma, assuming E[X] = 0, X ∈ [a, b] (w. prob. 1) and λ ∈ R: Lovász Pr B i ≥ (1 − xi ) > 0 where Pr[Bi ] ≤ xi · (1 − xj ),
(i,j)∈D
 2
λ (b − a)2

E eλX ≤ exp for xi ∈ [0, 1) for all i = 1, . . . , n and D the dependency graph.
 
8
If each Bi mutually indep. of all other events, exc. at most d,
1 1
V 
Pr[Bi ] ≤ p for all i = 1, . . . , n , then if ep(d + 1) ≤ 1 then Pr B i > 0.
  P
Kolmogorov Pr max |Sk | ≥ ε ≤ ε2
Var[Sn ] = ε2
Var[Xi ]
k i
where X1 , . . . , Xn are i.r.v., E[Xi ] = 0,
Var[Xi ] < ∞ for all i, Sk = ki=1 Xi and ε > 0.
P
✁✂ László Kozma · latest version: http://www.Lkozma.net/inequalities_cheat_sheet