Documente Academic
Documente Profesional
Documente Cultură
Appendix A
Summary of
Basic Formulae
Summary
Two sets of tables are provided for reference. The first records the definition, and
the first two moments of the most common probability distributions used in this
volume. The second records the basic elements of standard Bayesian inference
processes for a number of special cases. In particular, it records the appropriate
likelihood function, the sufficient statistics, the conjugate prior and corresponding posterior and predictive distributions, the reference prior and corresponding
reference posterior and predictive distributions.
A.1
PROBABILITY DISTRIBUTIONS
The first section of this Appendix consists of a set of tables which record the
notation, parameter range, variable range, definition, and first two moments of the
probability distributions (discrete and continuous, univariate and multivariate) used
in this volume.
428
x = 0, 1
p(x) = x (1 )1x
V [x] = (1 )
E[x] =
Bi(x | , n) Binomial (p. 115)
0 < < 1, n = 1, 2, . . .
n x
p(x) =
(1 )nx
x
x = 0, 1, . . . , n
E[x] = n
V [x] = n(1 )
x = 0, 1, . . . , n
c=
( + )
()()( + + n)
V [x] =
n ( + + n)
( + )2 ( + + 1)
N
N +M
x = a, a + 1, . . . , b
a = max(0, n M )
b = min(n, N )
1
N +M
c=
n
V [x] =
nN M N + M n
(N + M )2 N + M 1
429
x = 0, 1, 2, . . .
c = r
V [x] = r
1
2
x
p(x) = c
x!
E[x] =
x = 0, 1, 2, . . .
c = e
V [x] =
x = 0, 1, 2, . . .
( + x)
x
x!
( + )+x
E[x] =
()
V [x] =
1+
p(x) = c
c=
430
0<x<1
p(x) = c x1 (1 x)1
c=
E[x] =
( + )
()()
V [x] =
( + )2 ( + + 1)
a<x<b
c = (b a)1
E[x] = 12 (a + b)
V [x] =
1
12 (b
a)2
p(x) = c x
E[x] = 1
x>0
c=
()
V [x] = 2
x>0
c=
V [x] = 1/2
x>0
( + n)
c=
() (n)
V [x] =
2 (n2 + n( 1))
( 1)2 ( 2)
431
2 (x | ) = 2
>0
x>0
(1/2)/2
c=
(/2)
V [x] = 2
p(x) =
Pn i
2 (x | + 2i)
2
i=0
x>0
E[x] = +
V [x] = 2( + 2)
x>0
c=
()
V [x] =
2
( 1)2 ( 2)
1
2
x>0
(1/2)/2
c=
(/2)
V [x] =
2
( 2)2 ( 4)
( 1/2)
E[x] =
()
x>0
2
c=
()
V [x] =
E[x]2
1
432
> 0, > 0
p(x) = c x(+1)
E[x] =
>1
V [x] =
2
(1)2 (2)
>2
0 < x < 1
c =
V [x] = 2 ( + 1)2 ( + 2)1
< x < +
c = 1/2 (2)1/2
V [x] = 1
(+1)/2
p(x) = c 1 + 1 (x )2
< x < +
1/2
12 ( + 1)
c=
1
( 2 )
E[x] =
V [x] = 1 ( 2)1
F(x | , ) = F,
> 0, > 0
/21
p(x) = c
E[x] =
x
( + x)(+)/2
,
2
>2
x>0
12 ( + ) /2 /2
c=
( 12 )( 12 )
V [x] =
2 2 ( + 2) ,
>4
( 2)2 ( 4)
433
2
p(x) = 1 exp 1 (x ) 1 + exp 1 (x )
V [x] = 2 2 /3
E[x] =
k+1
x = (x1 , . . . , xk )
k
=1 x n
xi = 0, 1, 2, . . .
k
k
=1
, xk+1 = n
x
=1
=1
V [xi ] = ni (1 i )
E[xi ] = ni
C[xi , xj ] = ni j
[s] =
s
=1 (
E[xi ] = npi
i
pi = k+1
=1
+ 1)
x = (x1 , . . . , xk )
xi = 0, 1, 2, . . .
n
=1 xl n
c=
n!
k+1
=1
xk+1 = n
[n]
k
n+
=1
x
k+1
npi (1 pi )
n + k+1
=1
C[xi , xj ] =
npi pj
k+1
1 + =1
V [xi ] =
1+
=1
k+1
=1
434
E[xi ] = k+1
=1
=1
x = (x1 , . . . , xk )
k
0 < xi < 1,
=1 x 1
k+1
)
(
c = k+1=1
=1 ( )
C[xi , xj ] =
E[xi ]E[xj ]
1 + k+1
=1
V [y] = 2
= (1 , . . . , k ) k
symmetric positive-definite
p(x) = c exp 12 (x )t (x )
c = | | 1/2 (2)k/2
E[x] =
V [x] = 1
Pa2 (x, y | , 0 , 1 )
x = (x1 , . . . , xk ) k
(0 , 1 ) 2 , 0 < 1 , > 0
p(x, y) = c (y x)(+2)
0 1
1 0
E[x] =
E[y] =
1
1
435
(x, y) = (x1 , . . . , xk , y)
< xi < , y > 0
p(x, y) = Nk (x | , y) Ga(y | , )
E[x, y] = (, 1 ),
V [x] = ( 1)1 1 ,
p(x) = Stk (x | , 1 , 2)
V [y] = 2
p(y) = Ga(y | , )
x = (x1 , . . . , xk )
< xi < +
y symmetric positive-definite
p(x, y) = Nk (x | , y) Wik (y | , )
E[x, y] = {, 1 }
V [x] = ( 1)1 1
p(x) = Stk (x | , 1 , 2)
p(y) = Wik (y | , )
( 12 )()k/2
E[x] = ,
V [x] = 1 ( 2)1
x symmetric positive-definite
c = k
E[x] = ,
E[x ] = (
k+1 1
2 )
k(k1)/4 ||
=1
( 12 (2 + 1 ))
436
A.2
INFERENTIAL PROCESSES
The second section of this Appendix records the basic elements of the Bayesian
learning processes for many commonly used statistical models.
For each of these models, we provide, in separate sections of the table, the
following: the sufficient statistic and its sampling distribution; the conjugate family, the conjugate prior predictives for a single observable and for the sufficient
statistic, the conjugate posterior and the conjugate posterior predictive for a single
observable.
When clearly defined, we also provide, in a final section, the reference prior
and the corresponding reference posterior and posterior predictive for a single
observable. In the case of uniparameter models this can always be done. We recall,
however, from Section 5.2.4 that, in multiparameter problems, the reference prior
is only defined relative to an ordered parametrisation. In the univariate normal
model (Example 5.17), the reference prior for (, ) happens to be the same as that
for (, ), namely (, ) = (, ) 1 , and we provide the corresponding
reference posteriors for and , together with the reference predictive distribution
for a future observation.
In the multinomial, multivariate normal and linear regression models, however, there are very many different reference priors, corresponding to different
inference problems, and specified by different ordered parametrisations. These are
not reproduced in this Appendix.
Bernoulli model
z = {x1 , . . . , xn },
xi {0, 1}
p(xi | ) = Br(xi | ),
0<<1
n
t(z) = r = i=1 xi
p(r | ) = Bi(r | , n)
p() = Be( | , )
p(x) = Bb(r | , , 1)
p(r) = Bb(r | , , n)
p( | z) = Be( | + r, + n r)
p(x | z) = Bb(x | + r, + n r, 1)
() = Be( | 12 , 12 )
( | z) = Be( | 12 + r, 12 + n r)
(x | z) = Bb(x | 12 + r, 12 + n r, 1)
Negative-Binomial model
z = (x1 , . . . , xn ),
xi = 0, 1, 2, . . .
0<<1
p(xi | ) = Nb(xi | , r),
n
t(z) = s = i=1 xi
p(s | ) = Nb(s | , nr)
p() = Be( | , )
p(x) = Nbb(x | , , r)
p(s) = Nbb(s | , , nr)
p( | z) = Be( | + nr, + s)
p(x | z) = Nbb(x | + nr, + s, r)
() 1 (1 )1/2
( | z) = Be( | nr, s + 12 )
(x | z) = Nbb(x | nr, s + 12 , r)
437
438
Exponential Model
z = {x1 , . . . , xn },
0 < xi <
p(xi | ) = Ex(xi | ),
>0
t(z) = t = ni=1 xi
p(t | ) = Ga(t | n, )
p() = Ga( | , )
p(x) = Gg(x | , , 1)
p(t) = Gg(t | , , n)
p( | z) = Ga( | + n, + t)
p(x | z) = Gg(x | + n, + t, 1)
() 1
( | z) = Ga( | n, t)
(x | z) = Gg(x | n, t, 1)
Uniform Model
z = {x1 , . . . , xn },
0 < xi <
>0
p(xi | ) = Un(xi | 0, ),
t(z) = t = max{x1 , . . . , xn }
p(t | ) = Ip(t | n, 1 )
p() = Pa( | , )
p(x) = +1
Un(x | 0, ), if x ,
1
+n Ip(t | n, ),
1
+1 Pa(x | , ),
n
+n Pa(t | , ),
if x >
p(t) =
if t ,
if t >
p( | z) = Pa( | + n, n ), n = max{, t}
+n
1
p(x | z) = +n+1
Un(x | 0, n ), if x n , +n+1
Pa(x | , n ), if x > n
() 1
( | z) = Pa( | n, t)
n
(x | z) = n+1
Un(x | 0, t), if x t,
1
n+1 Pa(x | n, t),
if x > t
),
=
1
p( | z) = N(
|
n
n
n
n (0 0 + nx)
1
p(x | z) = N x | n , n (n + )
() = constant
( | z) = N( | x
, n)
(x | z) = N(x | x
, n(n + 1)1 )
439
440
441
k
r n
k =1 i
i=1 i 1
p( | z) = Dik | 1 + r1 , . . . , k + rk , k+1 + n k=1 r
p(x | z) = Mdk x | 1 + r1 , . . . , k + rk , k+1 + n k=1 r , n
442
Linear Regression
z = (y, X), y = (y1 , . . . , yn )t n , xi = (xi1 , . . . , xik ) k , X = (xij )
p(y | X, , ) = Nn (y | X, I n ),
k , > 0
t(z) = (X t X, X t y)
p(, ) = Ng(, | 0 , n0 , , ) = Nk ( | 0 , n0 ) Ga( | , )
p() = Stk ( | 0 , n0 1 , 2)
p() = Ga( | , )
p(y | x) = St y | x 0 , f (x) 1 , 2
f (x) = 1 x(xt x + n0 )1 xt ,
p( | z) = Stk | n , (n0 + X t X)( + 12 n)n1 , 2 + n ,
n = (n0 + X t X)1 (n0 0 + X t y),
n = + 12 (y X n )t y + 12 ( 0 n )t n0 0
p( | z) = Ga( | + 12 n, n )
p(y | x, z) = St(y | x n , fn (x)( + 12 n)n1 , 2 + n),
fn (x) = 1 x(xt x + n0 + X t X)1 xt
(, ) = (, ) (k+1)/2 (for all reorderings of the j )
n , 1 X t X(n k)1 , n k ,
( | z) = Stk |
n
2
n = (X t X)1 X t y,
n )t y
n = 1 (y X
2
( | z) = Ga( | 12 (n k), n )
n , 1 fn (x)(n k)1 , n k),
(y | x, z) = St(y | x
n
2
fn (x) = 1 x(xt x + X t X)1 xt