Formula Collection February 10, 2012

_
r
k
_
=
_
r 1
k 1
_
+
_
r 1
k
_
; (b)
_
r
k
_
=
r
k
_
r 1
k 1
_
=
r k +1
k
_
r
k 1
_
=
r
r k
_
r 1
k
_
. (1.1)
_
r
k
_
=
_
r
r k
_
; (b)
_
r
k, mk, r m
_
=
_
r
m
__
m
k
_
=
_
r
k
__
r k
mk
_
=
_
r
mk
__
r m+k
k
_
, 0 k r Z. (1.2)
_
r
k
_
(1)
k
=
_
r +k 1
k
_
; (b)
_
k +1/2
n
_
=
_
4
n
_
2n
n
_
(1)
n+k1
_
n
k+1
__
2n
2k+2
_
1
k < n
4
n
_
2n
n
__
2k+1
2n
__
k
n
_
1
k n
(1.3)
n

k=0
_
r +k
r
_
=
_
r +n+1
n
_
; (b)
n

k=m
_
r
k
_
(1)
k
=
_
r 1
m1
_
(1)
m
+
_
r 1
n
_
(1)
n
; (c)

k0
_
k
nk
_
= F
n+1
. (1.4)
m

k=0
_
k +n
t
_
=
_
m+n+1
t +1
_

_
n
t +1
_
(b)
n

k=m
_
k
m
_
=
_
n+1
m+1
_
; (c)
m

k=0
k
_
k +n
k
_
= (n+1)
_
m+n+1
m1
_
. (1.5)
n

k=1
k! k = (n+1)! 1; (b)
n

k=1
2k(2k 1)!! = (2n+1)!! 1; (c)
n

k=0
(2k +1)(2k)!! = (2n+2)!! 2. (1.6)
n

k=a
k
m
=
k
m+1
m+1

n+1
k=a
, m =1, a 0; (b)

k
_
n
k
_
2
=
_
2n
n
_
; (c)

k
k
_
n
k
_
2
= (2n1)
_
2n2
n1
_
; (1.7)
n

k=0
_
x +k
k
_
1
=
x
x 1
_
1
_
x +n
n+1
_
1_
; (b)
n

k=0
(1)
k
_
x
k
_
1
=
x +1
x +2
+
x n
x +2
(1)
n
_
x
n+1
_ . (1.8)
n

k=0
_
k +u
k
_
z
k
=
1
u!
d
u
dz
u
z
u
1z
n+1
1z
=
1z
n+1

u
i=0
_
i+n
i
_
(i z)
i
(1z)
u+1
, u N; (b)
n

k=0
_
1/2
k
_
(1)
k
=
_
1/2
n
_
(1)
n
. (1.9)
m

k=0
_
k +nm1
k
_
z
k
=
m

k=0
_
n
k
_
z
k
(1z)
mk
,
m n N;
s =

1+4z.
(b)
n

k=0
_
nk
k
_
z
k
=
(1+s)
n+1
(1s)
n+1
2
n+1
s
, (1.10)

k0
_
k +u
k
_
z
k
=
1
(1z)
u+1
; (b)

k0
_
s +k
u
_
z
k
=
z
us
(1z)
u+1
[u < s]
su1

k=0
_
u+k
u
_
z
k(su)
u, s N; (1.11)
n

k=0
_
x
k
_
k k!
x
k+1
= 1
_
x
n+1
_
(n+1)!
x
n+1
; (b)

k1
x
k
k k!
= E
i
(x) lnx =
e
x
x
_
1+O
_
x
1
_
_
. (1.12)

k0
_
r tk
k
_
z
k
=
x
r+1
(t +1)x t
, (b)

k0
_
r tk
k
_
r
r tk
z
k
= x
r
, where z = x
t+1
x
t
, (1.13)
and the solution is chosen be regular in z {|z| <|t
t
(t +1)
(t+1)
|}. This is also the region of convergence.
n

k=1
1
k
_
r +nk
r
_
=
_
r +n
n
_
_
H
n+r
H
r
_
. (b)
n

k=0
_
n+k
k
_
2
k
= 2
n
; (c)
n

k=0
_
2k
k
_
4
k
= (2n+1)
_
n1/2
n
_
; (1.14)

k1
_
n
k
_
(1)
k+1
k
(x +ky)
r
= x
r
H
n
+rx
r1
y, 0 r n. (b)
n

j=0
_
n
j
_
(1)
j
x + j
=
1
x
_
n+x
n
_, x / 0, 1, . . . , n. (1.15)

k
_
n
k
_
(1)
nk
k
n
= n! (b)

k
_
n
k
_
(1)
nk
k
n+1
=
1
/
2
n(n+1)! (c)

k
_
n
k
_
(tk +r)
k1
(t(nk) +s)
nk
=
(tn+r +s)
n
r
(1.16)
c
m,n

k
_
n
k
_
(a+bk)
m
(1)
k
= n!(1)
n

i
_
m
i
_
_
i
n
_
a
mi
b
i
[= 0 | m < n]; (b)

mn
needs |a+bk|<1, k[n]
c
m,n
=
1/(1a)
_
n(1a)/b
n
_ :=
(1, ab)
(1a)(n+1)
(1.17)
r1

k=0
_
n+1
k
_
(r k)
n
(1)
k
=

k0
_
nk
r 1
_
_
n
k
_
(1)
k+n+r+1
= A(n, r) =

n
r1
_
Eulerian #s. (1.18)

k0
_
2k
k
_
z
k
=
1
s
, (b)

k0
_
2k
k
_
kz
k
=
2z
s
3
, (c)

k0
_
2k
k
_
z
k
k +1
=
1s
2z
, (d)

k1
_
2k
k
_
z
k
k
= log
1s
z(1+s)
, s =

14z. |z| <

1
4
(1.19)

k0
z
k
k/2

r=0
_
k
r
_
x
r
=
1/2
1z(1+x)
_
1+
12xz

14xz
2
_
When bisected: (1.20)

l0
z
l
l

r=0
_
2l +1
r
_
x
r
=
1/2
1z(1+x)
2
_
1+x +
1x

14xz
_
; (b)

l0
z
l
l

r=0
_
2l
r
_
x
r
=
1/2
1z(1+x)
2
_
1+
12x(1+x)z

14xz
_
(1.21)
1
2Hofri: formulae collection 2
(n+1) = n! =

2n
n+1/2
e
n
_
1+
1
12n
+
1
288n
2
+O
_
n
3
_
_
n!! = 2
n/2
(1+n/2)

n/2n/2
.
_
2k +1
2
_
=

(2k)!
4
k
k!
(2.1)
_
n1/2
n
_
= (1)
n
_
1/2
n
_
= 4
n
_
2n
n
_

n
;
(n+a)
(n+b)
= n
ab
_
1+
(ab)(a+b1)
2n
+O
_
n
2
_
_
(2.2)
_
n
k
_
=
_
n
2k(nk)
_
1/2 _
n
k
_
k
_
n
nk
_
nk

_
1+
1
12
_
1
n

1
k

1
nk
_
+
1
288
_
1
n

1
k

1
nk
_
2
+O
_
1
kn(nk)
_
_
(2.3)

k
_
r
m+k
__
s
nk
_
=
_
r +s
m+n
_

k
_
r
m+k
__
s
n+k
_
=
_
r +s
r m+n
_
r Z
0
, m, n Z. (Vandermondes conv.) (3.1)
r

k=0
_
r k
m
__
s +k
n
_
=
_
r +s +1
m+n+1
_
n s 0, r, m 0.

k
_
l
m+k
__
s +k
n
_
(1)
k
= (1)
l+m
_
s m
nl
_
l 0. (3.2)

kr
_
r k
m
__
s
k t
_
(1)
k
= (1)
r+m
_
s m1
r t m
_
r, m Z
0
, t Z =
p,hR
+

k
_
p+hk
mk
__
p
k
_
(1)
k
=
_
h
m
_
, (3.3)

k
_
n+
k
__
n+
nk
_
t
k
= (t 1)
n
P
(,)
n
_
t +1
t 1
_
. P-Jacobi polynomial, G&R 8.96. (3.4)
r/2

k=0
_
n
k
__
n+r 2k 1
n1
_
(1)
k
=
_
n
r
_
;
r/2

k=0
_
n+r 1
k
__
2n+2r 2k 2
2n+r 1
_
(1)
k
= 2
r1
_
n+r 1
n
_
; n, r N; (3.5)

k
_
m
k
_
_
n
k
_ =
n+1
nm+1
n m.
t

k=0
_
t+vrs
tkr
_
_
t+v
tk
_ =
t+v+1
r+s+1
_
r+s
r
_ , s v; t +v r +s;

k
_
n
k
__
r sk
n
_
(1)
k
= s
n
r, s R (3.6)

k
_
r +tk
k
__
s tk
nk
_
r
r +tk
=
_
r +s
n
_
;

k
_
n
k
_
(r +tk)
k1
(s tk)
nk
=
(r +s)
n
r
r, s R (3.7)

k0
_
r +tk
k
__
s tk
nk
_
=

k0
_
pk
nk
_
t
k
=
_

_
t
p+1
(t 1)
np1
integer p < n
(1t
n+1
)/(1t) p = n
no closed form when p > n or when p / N
p r +s (3.8)

k0
_
m
k
__
nk
nk l
_
(t 1)
k
=

k0
_
m
k
__
nm
nk l
_
t
k
l, m, n Z
0
. (3.9)

k
_
1/2
k
__
1/2
n+1k
_
(1)
k
= 2
n
_
n
n/2
_
.

k
_
2n
k
__
2n+2
k
_
(1)
k
=
(1)
n
n+1
_
2n
n
_
. (3.10)
(1)
k+1
2
2k1

i0
_
1/2
i
_
1
2i +1
_
1/2
k i 1
_
1
2k 2i 1
=
1
2
2k+1

i0
_
i 1/2
i
_
1
i +1/2
_
k i 3/2
k i 1
_
1
k i 1/2
=
1
k
2
_
2k
k
_ (3.11)

k
_
x +1
k
__
x +a1
k +a1
__
x k
nk
_
=
_
x +a1
n+a1
__
x +n+a
n
_

k
_
x +y +k
k
__
y
ak
__
x
bk
_
=
_
x +a
b
__
y +b
a
_
(3.12)
n

k=0
(a+rk)q
k
=
a(a+nr)q
n+1
1q
+
rq(1q
n
)
(1q)
2
;
n

k=0
kq
k
=
q
(1q)
2
_
1(n+1)q
n
+nq
n+1

(4.1)
n

k=0
k
2
q
k
=
q
(1q)
3
_
1+q(n+1)
2
q
n
+
_
2n
2
+2n1
_
q
n+1
n
2
q
n+2
_
; S
(n)
r
(q)
n

k=1
k
r
q
k
: S
(n)
r+1
(q) = qD
q
S
(n)
r
(q). (4.2)
n

k=0
k(1)
k
= (1)
n
_
n
2
_
;
n

k=0
k
2
(1)
k
=
(1)
n
2
n(n+1);
n

k=a
k
m
=
k
m+1
m+1

n+1
k=a
, m =1, a 0; (4.3)
n

k=0
k
2
q
k
=
q
2
(1q)
3
_
2n(n+1)q
n1
+2
_
n
2
1
_
q
n
n(n1)q
n+1
_
; (b)T
r
(q) =

k
k
r
q
k
=
r!q
r
(1q)
r+1
. (4.4)
n

k=1
H
k
= (n+1)H
n
n = (n+1)(H
n+1
1) (b)
n

k=1
_
k
m
_
m0
H
k
=
_
n+1
m+1
__
H
n+1

1
m+1
_
, (c)
n

k=1
H
k
k
=
1
/
2
_
H
2
n
+H
(2)
n
_
(4.5)
n

k=1
H
k
k +1
=
1
2
_
H
2
n+1
H
(2)
n+1
_
; (b)
n

k=2
H
k
k 1
=
1
2
_
H
2
n1
+H
(2)
n1
_
+1
1
n
; (c)
n

k=1
H
2
k
= (n+1)H
2
n
(2n+1)H
n
+2n. (4.6)
Hofri: formulae collection . . . 33
2n

k=1
(1)
k
H
k
= H
n
/2; (b)
n

k=1
H
k
H
n+1k
= (n+2)(H
2
n+1
H
(2)
n+1
) 2(n+1)H
n
+2n. (4.7)
n

k=0
k
m
=
1
1+m
m

k=0
_
m+1
k
_
(n+1)
mk+1
B
k
=
B
m+1
(n+1) B
m+1
1+m
=
n
m+1
m+1
_
1+O
_
n
1
__
. (4.8)
B
k
, B
k
() are Bernoulli numbers, 6.12. and polynomials. (s, m) is the Hurwitz function.

n1
n

i=1
q
b+i
= qe
q

k0
(q)
k
k!
1
b+k +1
,

n1
n
n

i=1
q
b+i
= qe
q

k0
(q)
k
k!
q+k +1
b+k +1
, b / N (5.1)

n1
n
2
n

i=1
q
b+i
= qe
q

k0
(q)
k
k!

q+(q+k +1)
2
b+k +1
, b / N If b = 0 the sum is e
q
(1+q)q. (5.2)
n

k=l
k=0
1
k l
_
(1)
l
_
n
l
_
+(1)
k
_
n
k
__
= 0, 0 l n;

n0
n

i=1
q
r +i
=
r!
q
r
_
e
q
e
r1
(q)
_
, r N; (5.3)

j0
(qz)
j
(z +1)(z +q) (z +q
j
)
= 1, q, z C, |q| > 1 (or |q| = 1, |z| < 1); (5.4)
n

j=1
a
k
j

n
i=1
i=j
(a
j
a
i
)
=
_

_
0 k < n1
1 k = n1
h
kn+1
(aaa) k n

distinct a
i
C;
h
r
(aaa) =
i
j
0
i
j
=r

n
j=1
a
i
j
j
symm. poly
=
n

j=1
n

k=j
k=1
z a
j
a
k
a
j
= 1. (5.5)

k
(t
1
, . . . t
n
) = [z
nk
](z +t
1
) (z +t
n
), the kth elementary symmetric polynomial in the n variables t
j
, 1 k n.
Let S
m

n
i=1
t
m
i
, then S
m
=

i
1
+2i
2
++ni
n
=m
(1)
i
2
+i
4
+i
6
+
(i
1
+i
2
+ +i
n
1)!m
i
1
!i
2!
i
n
!

i
1
1

i
2
2

i
n
n
, Waring formula (5.6)

r0
(uv)
r
z
(
r
2
)

r
i=0
(1vz
i
)
=

i0
v
i
i1

j=0
_
1+uz
j
_
; viz. (5.8)(a). The case k = n1 in (5.5) is Good identity.. (5.7)

n0
u
n

n
i=1
(1q
i
)
=
1

i0
(1uq
i
)
,

r0
u
r
z
(
r
2
)

r
i=1
(1z
i
)
(a)
= 1+u

i0
z
i
i1

j=0
_
1+uz
j
_
=

k0
_
1+uz
k
_
, Euler identities (5.8)
(uz)

(z)

=

n0
(u)
n
z
n
(q)
n
, with (a)
n
= (a; q)
n

n1

i=0
_
1aq
i
_
; (a; q)
0
= 1; (a; q)
t
=
(a; q)

(aq
t
; q)

,
Cauchy identity
qPochhammer symbol
(5.9)

n0
(a)
n
(b)
n
(c)
n
(q)
n
t
n
=
(b)

(at)

(c)

(t)

k0
(c/b)
k
(t)
k
(at)
k
(q)
k
b
k
Heines transformation (5.10)
n

k=1
a
k
b
k
= a
k
b
k

n+1
1

n

k=1
_
a
k
_
b
k+1
summation by parts. (6.1)

1i, jn
min
_
x
i
, x
j
_
= 2
_
n

i=1
ix
(i)

i=1
x
i
_
, where x
(1)
x
(2)
x
(n)
. (6.2)
M

m=1
N

n=1
min(m, n) =
1
6
N(N+1)(3MN+1).
M

m=1
N

n=1
max(m, n) =
1
6
N
_
N
21
_
+
1
2
MN(M+1). M N (6.3)
_
e
t
t
n
dt =
e
t

n+1
n!
n

i=0
(t)
i
i!
;
_

t=0
e
t
(a+t)
j
dt =
j!

j+1
j

i=0
(a)
i
i!
; (6.4)
_
1
t=0
t
z1
(1t)
w1
dt =
_

u=0
u
z1
(1+u)
w+z
du = B(z, w) =
(z)(w)
(w+z)
, (6.5)
_

x
2 j
e
ax
2
dx =
1
a
(2 j+1)/2

_
j +
1
2
_
=

(2j)!

a(4a)
j
j!
(6.6)
_

e
ax
2
dx =
_
/a, (b)
_

e
ax
2
+bx
dx =
_
/ae
b
2
/4a
, (c)
_

0
e
ax
2
+bx
dx =
_
/ae
b
2
/4a

_
b/

2a
_
. (6.7)
_

0
xe
ax
2
+bx
dx =
1
2a
+
b
a
_
/ae
b
2
/4a

_
b/

2a
_
, (b)
_

xe
ax
2
+bx
dx =
b
a
_
/ae
b
2
/4a
(6.8)
_

0
x
2k
e
ax
2
dx =
(2k)!
_
/a
2(4a)
k
k!
, (b)
_

0
x
2k+1
e
ax
2
dx =
k!
2a
k+1
see Eq.6.15 (6.9)
4Hofri: formulae collection 4
n

k=1
f (k)=
_
n
1
f (x) +1/2[ f (n) + f (1)] +
m

j=1
B
2 j
(2j)!
_
f
[2 j1]
(n) f
[2 j1]
(1)
_
+
1
(2m+1)!
_
n
1
B
2m+1
({x}) f
[2m+1]
(x)dx. (6.10)
_
n
2
_
+
_
n
2
_
= n :
_
n+1
2
_
=
_
n
2
_
:
_
n1
2
_
=
_
n
2
_
1 (6.11)
B
k0
: 1,
1
/
2
, 1/6, (even) 1/30, 1/42, 1/30, 5/66, 691/2730, 7/6, 3617/510, 43867/798, 174611/330 (6.12)
m

k=1
k
s
=(s) (s, m+1); v(v +1, m+1)

j0
v
j
B
j
j!m
j+v
=
1
m
v

v
2m
v+1
+
v(v +1)
12m
v+2
+ ; m (see (6.12)) (6.13)
ln(1+x) =

r1
(x)
r
r
=

r1
1
r
x
r
_
1x
r
_
, |x| < 1. ln(x)
(x1/2)
=

k1
(x 1)
k
kx
k
(x>0)
=

k0
1
2k +1
_
x
2
1
x
2
+1
_
2k+1
. (6.14)
1(x)
(x)
x
_
1x
2
+3x
4

_
(A&S, 26.2.12), erf(x) = 2
_
x

2
_
1. (6.15)
INEQUALITIES
Markov:
E|X|
r
a
r
a.s. sup|X|
r
P{|X| a} E
__
|X|
a
_
n
_
r, n 0. (7.1)
This is the best possible when the only information is that X is non-negative and with nite mean. When it is also known that X is
unimodal about M, then P[|X M| a] <
_
n
n+1
_
n
E
_

X M
a

n
_
.
Inequality (7.1) for integer rvs and n = 1 becomes: P(X > 0) E[X]. also for any rv P(X = 0)
2
/E[X
2
]. (7.2)
Bienaym` eChebyshev: P{|X| t} t
2
E[X
2
] same as the RHS of (7.1) for n = 2. (7.3)
Chebyshev Cantelli: P{X t}

2

2
+t
2
(7.4)
Both (7.1) and (7.3) are special cases of the following, more general inequality (Lo eve, Probability Theory Vol I, p. 159): Let X be an
arbitrary r.v. and g on R be a nonnegative Borel Function. If g is even and nondecreasing on [0, +) then, for every a 0
Lo eve:
E[g(X)] g(a)
a.s. supg(X)
P[|X| a]
E[g(X)]
g(a)
(7.5)
If g is nondecreasing on R, then the middle term is replaced by P[X a], where a is an arbitrary number.
Let X
1
, X
2
, . . . , X
n
be independent rvs with E X
2
i
<, and S
k
the sum of the rst k X
i
s. Then, for any > 0
Kolmogorov: P
_
max
1kn

S
k
E[S
k
]

2
n

2

2
n
V[S
n
]. (7.6)
This also holds when S
n
is a martingale. For n = 1 (7.6) reduces (7.3).
Let X
i
, 1 i n be i.i.d. bounded rvs, 0 X
i
1, = E[S =X
i
]/n, then, for 0 <t < 1:
Hoeffding: P[Sn nt]
_
_

+t
_
+t
_
1
1 t
_
1t
_
n
exp
_
nt
2
g()
_
exp
_
2nt
2
_
, (7.7)
where
g() =
_
1
12
log(
1

) 0 < <
1
2
1
2(1)
1
2
< 1
Note : g() 2, .
For t > 1 the rst bound in the right-hand-side in (7.7) is 0, and for t = 1 it equals
n
. (Hoeffding, J. Am. Stat. Ass. 58, 1330,
1963). The right-most inequality is often named after Chernoff.
The combinatorial Bonferroni inequalities are easily adapted probability measures (Comtet, pp. 1923). Let A = {A
i
, 1 i n} be a
set of events in some probability space , associated with a measure P, and dene the probabilities S
k
as the sum of probabilities of
occurrence of the intersections of all k-strong subsets of these events: S
0
P(), S
k
=

1i
1
<i
2
<<i
k
n
P
_
A
i
1
A
i
2
A
i
k
_
. Then the sum

n
k=1
(1)
k1
S
k
satises the alternating inequalities:
Bonferroni: (1)
k
_
P
_
n
_
i=1
A
i
_
+
k

j=1
(1)
j
S
j
_
0 (1)
k+1
_
P
_
n

i=1
A
i
_
+
k

j=0
(1)
j+1
S
j
_
0 1 k n (7.8)
For k = 1 the above reduces Booles inequality.
Hofri: formulae collection . . . 55
Let
i
be moments of X, and L(s) its LST. L(s) is absolutely monotone, and
Laplace-Stieltjes Transform:
2n1

i=0

i
(s)
i
i!
L(s)
2n

i=0

i
(s)
i
i!
n 1. (7.9)
Jensen: E[u(X)] u(E[X]) when u() is a convex function. (7.10)
Chebyshev: n
n

k=1
a
k
b
k

_
n

k=1
a
k
__
n

k=1
b
k
_
where {a
i
} and {b
i
} are monotonic, both increasing or decreasing, (7.11)
If the senses of monotonicity are opposite, reverse the inequality sign.
H older:
n

k=1

a
k
b
k

_
n

k=1

a
k

p
_
1/p
_
n

k=1

b
k

q
_
1/q
for p > 1, q > 1, and p
1
+q
1
= 1. (7.12)
The Cauchy-Bunyakovsky-Schwartz (CBS) inequality is obtained for p = q = 2. The analogue holds for integrals.
The CBS inequality also leads the following (in Vitter & Krishnan, JACM 43, #5 p.785):
Amit-Miller:
_
n

k=1

p
k
r
k

_
2
2
n

k=1
p
k
ln(p
k
/r
k
), for any two n term PMFs ppp and rrr (7.13)
Gibbs:
n

k=1
p
k
log(p
k
/r
k
) 0, for any two n term PMFs ppp and rrr (with equality if the PMFs coincide). (7.14)
Minkowski-I:
_
n

k=1
_
a
k
+b
k
_
p
_
1/p

_
n

k=1
a
p
k
_
1/p
+
_
n

k=1
b
p
k
_
1/p
. for p > 1 and a
k
, b
k
> 0, k. (7.15)
The analogue holds for integrals.
Minkowski-II:
n

k=1
_
1+a
k
_

_
_
1+
_
n

k=1
a
k
_
1/n
_
_
n
. for a
k
0 for all k and integer n. (7.16)
Carleman:
n

k=1
_
a
1
a
2
a
k
_
1/k
e
n

k=1
a
k
for a
k
0 for all k, and an integer n. (7.17)
On the interval [a, b] let f () be real and continuous and let there be a family (not necessarily nite) of real and continuous orthonormal
functions, {
r
()}, then
Bessel:

r
_
_
b
a
f (x)
r
(x)dx
_
2

_
b
a
f
2
(x)dx, The conditions on f , which may be complex, can be relaxed. (7.18)
Hilbert:

m,n
a
m
b
n
m+n
<

sin(/p)
A
1/p
B
1/q
, when p, q > 1, 1/p+1/q = 1,
m
a
p
m
A,
n
b
q
n
B. (7.19)
unless one of the series vanishes identically. (More on this in Hardy, Littlewood and P olya, Chapter IX).
interpolation:
n
n+1
m
m1
_
n1
n
_
nm+1
1, 1 < m < n. (7.20)
The wedge inequality: |(ab) (c d)| |ac| |bd| a, b, c, d R; : selection of the maximum : (7.21)
The inequality holds also for functions instead of numbers, with | | replaced by any p-norm, p > 0.
The ratio inequality:
n

i=1
a
i
_ n

j=1
b
j
max
k
a
k
b
k
a
k
, b
k
R
+
, (7.22)
Let A, G and H denote the arithmetic, geometric and harmonic means of {a} ={a
i
, 1 i n}, given by
The Three Means: A =

a
i
/n, G = [

a
i
]
1/n
, H = [

a
1
i
/n]
1
. (7.23)
The standard relation A G H can be generalized as follows: Let M
t
{a} (
1
n
a
t
i
)
1/t
, then
Power-means (Arnon): t
1
>t
2
=M
t
1
{a} M
t
2
{a}. (7.24)
A, G, H correspond to t = 1, 0, 1. As t , M
t
{a} approaches max{a} and min{a} resp. The power-means ineq. can be further
generalized with weights: Let p a PMF, and E
p
t
{a} (p
i
a
t
i
)
1/t
, then
Weighted power-means: t
1
>t
2
=E
p
t
1
{a} E
p
t
2
{a}. (7.25)
Let N
t
{a} (a
t
i
)
1/t
, with all a
i
> 0, then
Norm monotonicity: t
1
>t
2
=N
t
1
{a} N
t
2
{a}. (7.26)
6Hofri: formulae collection 6
The Jackknife inequality (Efron & Stein Ann. Statist. 9, 586596, 1981).
Let {X
i
}, 1 i n be i.i.d and H a symmetric measurable function over n1-dimensional vectors.
Dene S
i
= H(X
1
, . . . , X
i1
, X
i+1
, . . . , X
n
), and S =
n
i=1
S
i
/n. Then
The Jackknife inequality: V
_
H(X
1
, . . . , X
n1
)

E
_
n

i=1
(S
i
S)
2
_
. (7.27)
Shachnai: 0 < q < 1, a
i
0,
n

i=1
a
i
q
i
i1

j=1
a
j

1+3q
2(1q)
n

i=1
a
2
i
q
i
(7.28)
Shown in A. Bar-Noy, M. Bellare, M. Halldorsson, H. Shachnai, T. Tamir, Inf. and Comput., 140, 1998, 183202.
4
n
2

n

_
2n
n
_

4
n

3n+1
; (b)
_
n
k
_
k

_
n
k
_

_
n
n
k
k
(nk)
nk
,
_
en
k
_
k
_
. (7.29)
Logarithm: lim
x
x
a
logx = 0, (b) lim
x0
x
a
logx = 0, a > 0 (c) lnx n(x
1/n
1), x > 0, n Z
+
(7.30)
x
x +1
< ln(1+x) < x, 0 = x >1, (b) x <ln(1x) <
x
1x
, 0 = x < 1, (c) lnx x 1, x > 0 (7.31)
| log(1+z)| log(1|z|). |z| < 1. (b)
_
1+
1
x
_
x
x > 0 is monotonic incr. (7.32)
Exponent: e

x
1x
< 1x < e
x
, x [0, 1), (b)
1
1x
> (x < 1) e
x
> 1+x, (c)
x
1+x
< 1e
x
< x, (x >1). (7.33)
1+x > e
x
1+x
x >1, (b) e
x
>
_
1+
x
y
_
y
> e
xy
x+y
x, y > 0 (c) |e
z
1| e
|z|
1 |z|e
|z|
all z. (7.34)
Let m
a
= E[|X EX|
a
], a R, then m
1/a
a
is a monotonic increasing function in a. See (7.24) (8.1)
_
1
p=0
F
r
p
dp = E(X
r
), where F
p
is the p fractile of the rv X : P
_
X F
p
_
= p. (8.2)

i1
_
u
t=0
[1G(ut)]g
i
(t)dt = G(u), g(t) =
dG(t)
dt
; 1(x)
(x)
x
_
1x
2
+3x
4

_
(A & S, 26.2.12). (8.3)
For the M/G/1 queue: PGF of customers in system, mean value and variance, where S(t) is the LST of the service time distribution and

r
is E(S
r
):
g(z) = p
0

S(z)(z 1)
z

S(z)
,

S(z) S( z), p
0
= 1E(S) = 1, E(X) = +

2

2
2(1)
, E(W) =

2
2(1)
(8.4)
E(X
2
) = E(X) +

3

3
+3
2

2
3(1)
+

4

2
2
2(1)
2
, W(s) =
s(1)
S(s) +s
.
For the M/G/ queue: if arrivals are in batches where the batch size has the PGF A(z), and F() is the service time CDF, the number in
queue has the Poisson distribution only when A(z) = z:
The Expected length of a bp and of the number of customers served in a bp are given by
E[B] =
e

, E[N] =e

, = A

(1),
_

t=0
(1A(F(t)))dt. (8.5)
When A(z) = z, = E[S], E[B] = (e

1)/, E[N] = e

. (Shanbhag, J. Appl. Prob. 3, 274279 (1966)).

Volume of an n-dimensional ball of radius R:
V
n
(R) = 2
n/2

n/2
R
n
/n!! =
n/2
R
n
/(n/2+1). A
n
(R) = dV
n
(R)/dR (8.6)
Moments of N(a,
2
) : E(X a)
2r
= (2r)!
2r
/2
r
(r +1). E|X a|
r
= 2
r/2
((r +1)/2)
r
/

, odd r (8.7)
The eigenvalues of an n-circulant with rst column c are given by
=

nWWWc, where W
jk
= exp
_
i
2
n
jk
_
. (8.8)
GENERIC EQUATIONS
The general linear differential equation of rst order:
y

(x) +h(x)y(x) = g(x), (9.1e)

with h() and g() known functions, and a prescribed boundary value y
0
= y(x
0
), has the general solution
y(x) = e

_
hdx
_
C+
_
ge
_
hdx
dx
_
, (9.1s)
Hofri: formulae collection . . . 77
where the constant C is determined from the boundary value.
The Bernoulli equation is y

+ p(x)y = q(x)y
n
; the substitution v = y
1n
linearizes it to v

= (n1)[p(x)v q(x)].
Homogeneous Ricatti equation with constant coefcients, aka logistic equation (a particular case of Bernoulli equation):
f

(x) = af
2
(x) +bf (x), f (0) =, where a and b are constants, has the solution (9.2e)
f (x) =
e
bx
1ab
1
(e
bx
1)
, (Goulden and Jackson, 1984). (9.2s)
The general linear rst order difference equation
a
t+1
=
t
+
t
a
t
; t m, with a
m
, {
k
}, {
k
} known, (9.3e)
has the solution
a
t
=
t1

j=m

j
t1

i=j+1

i
+a
m
t1

i=m

i
t m (9.3s)
Special cases: Constant and
a
t+1
= +a
t
; t m = a
t
=

1
+
tm
_
a
m

1
_
, t m. (9.3c)
Pure linear
t
, constant
t
:
a
i+1
= b+ica
i
, i 1, a
1
given (9.5e)
a
i
= (i 1)!c
i1
_
a
1
+b(e
i1
(1/c) 1)

, i 1. e
r
(u) the incomplete exponential function. (9.5s)
Functional equation: a(x) and b(x) given functions and |p| = 1.
f (x) = a(x) +b(x) f (px), A = lim
i
p
i
, f (A) known. (9.6e)
In the region where the following sums and products converge, the solution is given by
f (x) =

j0
a
_
p
j
x
_
j1

i=0
b
_
p
i
x
_
+ f (A)

i0
b
_
p
i
x
_
(9.6s)
T(n) = n(lgn)
q1
+2T(n/2), q N T(n) (n/q)(lgn)
q
. (9.7)